Prepared for submission to JHEP

A Traversable Wormhole Teleportation Protocol in

the SYK Model
arXiv:1911.07416v4 [hep-th] 8 Aug 2021

Ping Gaoa and Daniel Louis Jafferisb

a
Center for Theoretical Physics, Massachusetts Institute of Technology,
Cambridge, MA 02139, USA
b
Center for the Fundamental Laws of Nature, Harvard University,
Cambridge, MA 02138, USA

E-mail: pgao@mit.edu, jafferis@g.harvard.edu

Abstract: In this paper, we propose a concrete teleportation protocol in the SYK model
based on a particle traversing a wormhole. The required operations for the communication,
and insertion and extraction of the qubit, are all simple operators in terms of the basic
qubits. We determine the effectiveness of this protocol, and find a version achieves almost
perfect fidelity. Many features of semiclassical traversable wormholes are manifested in this
setup.
Contents

1 Introduction 1

2 A simple wormhole teleportation protocol 4

2.1 General setup 4
2.2 SYK model 9
2.3 Density matrix ρT R in SYK model 12

3 Computing the correlation functions in the SYK model 13

3.1 Stitching correlation function 13
3.1.1 Symmetries 13
3.1.2 Twist boundary condition 15
3.1.3 Smoothness and UV conditions 16
3.2 Large q solution 17
3.2.1 General solution 17
3.2.2 Translation invariant solution: µ = 0 case 18
3.2.3 Twist solution at large q 19

4 Effectiveness of protocol 21
4.1 Lorentzian correlation function 22
4.2 Semiclassical limit 23
4.3 Improved protocol with multiple fermion SWAP 29

5 Discussion 31
5.1 Interference regime 31
5.2 Partial coupling 32
5.3 Teleportation of multiple qubits 34
5.4 Classical simulation and separability 35

6 Conclusion 36

A Solution of σ 37

1 Introduction

Recent constructions of traversable wormholes [1, 2] provide a causal probe of the ER=EPR
[3] relation between entanglement and geometry. A pair of black holes in thermofield
double state have their interiors connected via a Einstein-Rosen bridge behind the horizon.
The Einstein-Rosen bridge is non-traversable so signals cannot get through the wormhole
although the black hole interiors are spatially connected. The basic mechanism found in

–1–
[1] is that the gravitational backreaction to quantum effects induced by generic couplings
between the exterior regions of the pair of black holes can render the wormhole traversable.
One of the most interesting consequences is that a region of connected spacetime which
would have been cloaked behind the horizon becomes visible to the external boundaries.
Many recent works have studied traversable wormholes in more general settings, e.g. [4–8].
It was conjectured by [1, 2] that sending quantum information through such a wormhole
is the gravitational description of quantum teleportation in the dual many-body system.
Let us denote the two black holes by l and r, which are initially in the thermofield double
state. The physical picture behind this teleportation is quite simple: the qubit that is
teleported from l to r is carried by a particle that travels through the dual wormhole.
For a qubit thrown into one side of the system at a sufficiently early time, one just
needs to wait long enough (of order the scrambling time) so that the qubit is sufficiently
mixed with the black hole microstates, then apply a simple coupling between l and r, and
finally wait for another scrambling time for the qubit to appear on the other side.
Many subsequent works developed this picture in more detail. In [9], a scrambling
process combined with teleportation protocol was compared with the traversable wormhole
protocol. Their protocol requires precursor operators with a much higher complexity than
the simple coupling between l and r. In [10], the traversable wormhole was understood as
a quantum channel. In [11], a nontrivial protocol was proposed to decode a qubit that has
been highly scrambled in a black hole background. However, not all of the features of the
protocol suggested by traversable wormholes are manifested in these versions, particularly
that fact that only simple operators are required. 1 Moreover, as we will discuss further
below, other effects can lead to teleportation associated to fully scrambled (essentially
random unitarity) dynamics, one sign of which is time inversion of the transmitted quantum
information.
In this paper, we propose a precise teleportation protocol using N qubit fermionic
systems that correspond to gravity in a nearly AdS2 black hole spacetime. We will deter-
mine the fidelity of teleportation by computing the mutual information between a reference
qubit that is initially in a Bell state with the injected qubit on the left and a final register
qubit on the right. We find an optimal method for encoding the qubit in terms of the basic
fermions that leads to perfect teleportation for large N in a certain limit on the parame-
ters describing the dynamics of fermionic system. More generally, we also investigate the
separability of the final state when the mutual information is sub-maximal, to determine
when the information transfer is truly quantum.
We examine this in detail in the SYK model [13, 14] with q-point random interactions
[15], which is solvable in large N limit. Our protocol invovles applying SWAP operations
constructed out of simple operators of the SYK model to insert the qubit into the black
hole from the left side and extract it on the right side, as well as a communication channel
between l and r that involves a coupling of only the basic fermions, of the type used in
[2, 16] (see Fig. 1). Recent works studying traversable wormholes using the SYK model
1
A very recent work [12] proposed a protocol of teleportation by size of operators, which in that sense
is similar to traversable wormholes.

–2–
can been found in e.g. [17–23].
One can construct Dirac fermions out of pairs of the basic Majoranas, and the tele-
ported qubits are encoded in those Dirac fermion occupation numbers. It is particularly
interesting to examine situations which have a clear dual description in terms of a particle
physically propagating through the wormhole. If our qubit was encoded in the occupation
number of a bulk fermion mode, then at low temperatures, it would be distinguishable from
the thermal atmosphere. Note that at high temperatures, there is a significant probability
that a given mode is already occupied in the Hartle-Hawking state itself.
The bulk description of the Majorana SYK model is slightly more complicated at
the boundary, and the fundamental fermions have occupation number hN̂ i = 21 in the
thermal state at any temperature. Nevertheless, the correlation function between the
number operators in the left and right systems are small at low temperature, and the
desired behavior can still be seen. In particular, after SWAP the qubit into SYK, the
occupation number of fundamental fermions will change according to the qubit.
We find the associated solution in the classical large N limit in terms of the collective
bilocal variables G and Σ. To solve the classical equations we further specialize to the
large q limit where they simplify. Even at low temperatures, our analysis goes beyond the
leading Schwarzian approximation.
The effect of the instantaneous interaction required for teleportation can be described
exactly in terms of its action on the basic Majoranas. This results in a non-trivial trans-
formation of the bilocal field G that we determine. Thus in the large N limit, the problem
reduces to finding a solution with that twist boundary condition across the moment of
interaction.
Features of the full G-Σ solution beyond the Schwarzian mode are important in under-
standing teleportation. In particular, in the Schwarzian approximation, the leading order
1/N classical solution for the collective reparametrization field due to the interaction would
lead to a pole in the response measured by the occupation number on the right, whereas
Fermi statistics demands that this number is bounded by 1. We will find a consistent result
from classical G-Σ solution, which allows us to determine the fidelity of the teleportation
protocol.
An exact bulk dual of the SYK model is not yet known, but we will often refer to
the gravity description for intuition and to interpret our results. At low temperatures in
the large N limit, the correlation functions of fermions in SYK agree with free fields in
the two dimensional bulk spacetime, corrected at leading order in 1/N by exchanges of the
Schwarzian soft mode. This approximation, which is equivalent to bulk fermions minimally
coupled to Jackiw-Teitelboim gravity, suffices to match the physics of traversable wormholes
in the appropriate regime.
On the gravity side, stringy corrections lead to inelastic effects in the out of time order
correlators that describe quantum chaos [24, 25], and we will analogously refer to such
features in SYK in that way. The G-Σ boundary bi-local variables, however, have not yet
been recast in a manner that looks like a bulk string field theory.
We find that many features of semiclassical traversable wormholes are manifested in
this setup. In particular, the coupling between l and r can be rather general while main-

–3–
taining the fidelity of the protocol. Moreover, if we insert a sequence of qubits into the
system in an appropriate time window, these qubits will come out with the same time
ordering, as indicated by the smooth semiclassical geometry of a traversable wormhole.
This should be contrasted with what occurs in what one might call the fully scrambled
regime, when the evolution eiHt is effectively equivalent to a random unitary. Signals can
still reach the right, however they appear in reverse time order. We check that with our
protocol some quantum information can be sent in this case only at very high temperatures,
although the fidelity is submaximal.
At low temperatures in the gravity description, the fully scrambled regime corresponds
to a particle sent so early into the black hole that it has a destructive ultra high energy
collision with the negative energy squeezed state generated by the interaction protocol and
it does not smoothly pass through the wormhole. Nevertheless signals can be detected
on the right due to a quantum interference effect [2, 26]. Our protocol involves encoding
a qubit into one of the fermions, rather than simple signaling as measured by a left-right
causal propagator. It would be interesting to understand if there is a bulk description of the
partially successful teleportation we find at very high temperatures in the fully scrambled
regime.
We also find another regime in which qubits arrive on the right in inverse time order.
This appears at any temperature when the qubit is injected too late. In that case, the
traversing trajectory will be very close to ultimate horizon of the traverable wormhole, and
it is likely that quantum fluctuations of the geometry, or stringy effects related to the AdS
boundary in whatever is the exact dual of the SYK model are important. We leave further
exploration of that effect in the bulk description to future work.
An important feature of the traversable wormhole teleportation protocol is that the
qubit travels through the deep interior of the spacetime, and thus it is insensitive to single
boundary simple unitary operations after the qubit is sufficiently scrambled. Therefore this
is teleportation of a highly error protected qubit.
The structure of this paper is as follows. In Sec. 2, we introduce the protocol and
describe its detailed realization in the SYK model. To assess the effectiveness of the
protocol, we need to calculate various correlation functions, which we do in Sec. 3, in
the large q limit. In Sec. 4, we demonstrate the effectiveness of the protocol. In Sec.
5 we discuss four aspects of the protocol: the fully scrambled interference regime, more
general choices of coupling, teleportation of multiple qubits, and whether any quantum
information is transmitted in cases with submaximal fidelity. We conclude in Sec. 6.
Appendix A contains further details of the solution of the key correlation function.

2 A simple wormhole teleportation protocol

2.1 General setup

Consider 5 systems: R, Q, l, r, T. Let R, Q, T be single qubit systems, where R will play
the role of a reference, and information will transferred from Q to T. Let l and r be two
identical systems of N/2 qubits, where we will take N to be large.

–4–
Figure 1: The protocol via SWAP operators. Ut = e−iHt is the time evolution of SYK
system.

We will consider l and r to be physically realized as systems of N Majorana fermions,

pairs of which can be identified as Dirac fermions. Taking l and r together there are a total
of N qubits. Time evolution will be generated by the SYK model Hamiltonian acting on
these fermionic systems, as discussed further in the next section.
A Majorana fermion presentation of a qubit system can be constructed as follows 2
1 1
ψl2j−1 = √ (ZX)j−1 XX(II)N/2−j , ψl2j = √ (ZX)j−1 Y X(II)N/2−j (2.1)
2 2
1 1
ψr2j−1 = √ (ZX)j−1 IY (II)N/2−j , ψr2j = √ (ZX)j−1 IZ(II)N/2−j (2.2)
2 2
where X, Y, Z, I stand for the Pauli matrices and 2 by 2 identity, respectively. In the
above notation, each Majorana fermion is tensor product of N 2 by 2 matrices acting on
N distinct qubits in which odd numbered qubits are assigned to the l system and even
numbered qubits are assigned to the r system. For notational simplicity, we omit all
tensor product symbols and powers An should be interpreted in the tensor product sense
A ⊗ · · · ⊗ A.
Initially, RQ are in a maximally entangled state
1
|φ0 i = √ (|00i + |11i) (2.3)
2
l and r are in the thermofield double state
1 X −βEn /2
|tfdi = e |nnilr (2.4)
Z n

and T is in the state |0i. Success of teleportation from Q to T using the entanglement
resource of l and r is equivalent to producing maximal entanglement between R and T. 3
The traversable wormhole protocol is as follows (see see Fig. 1).
2
It would be interesting to study if these operators are easy to manipulate in some experimental real-
ization of the SYK model.
3
Maximal entanglement between R and T means that they are in one of four Bell states, which are
related to each other by a Pauli matrix operation.

–5–
 1. At time −t0 , apply a SWAP between Q and l such that the qubit is inserted into the
wormhole.

2. At time zero, apply an interaction U = eiµV with V = Ol Or to the state.

3. At time t, apply a SWAP between T and r such that the qubit is extracted from the
wormhole.

In the first step, we need to find an appropriate SWAP operator. Since l and r contain
Dirac fermions, we can pick one χl and its conjugate χ†l with

{χl , χ†l } = 1, χ2l = χ†2

l =0 (2.5)

and construct a SWAP operator as

!
χl χ†l χ†l
SQl = (2.6)
χl χ†l χl

Note that this is the ordinary SWAP operator, as can be seen from the Pauli matrix
representation for χl and χ†l :
! !
01 † 00
χl ' , χl ' (2.7)
00 10

In the final step of the protocol, we will use the same definition for the SWAP operator
ST r by replacing l with r in (2.6).
The second step can equivalently be replaced by measurement on the left together with
action of a unitarity operator on the right, so that only a classical channel communication
is required. The basic idea is that an interaction eiµOl Or =
P l iaµOr
a Pa e , where Pal is
the projector on to the eigensubspace of Ol with eigenvalue a. Whether the value of a is
actually measured, that is, classically recorded, on the left cannot affect the result on the
right [2].
Our protocol will involve a slight generalization, in which many operators are coupled,
iµ0 j Olj Orj
P
e , which is equivalent to a classical channel only when the various Olj Orj commute
with each other. In particular, the construction will involve coupling the Majoranas, which
can be written in terms of operations on the qubits using the above representation as
1 1
iψl2j−1 ψr2j−1 = − (XZ)j , iψl2j ψr2j = (Y Y )j (2.8)
2 2

where [· · · ]j ≡ (II)j−1 [· · · ](II)N/2−j means acting only on j-th left and right qubits.
It is straightforward to see that
2j−1 2j−1
h i
eiµ0 ·iψl ψr = Px+ e−iµ0 Z/2 + Px− eiµ0 Z/2 (2.9)
j
2j 2j
h i
eiµ0 ·iψl ψr = Py+ eiµ0 Y /2 + Py− e−iµ0 Y /2 (2.10)
j

–6–
where Px± = (I ± X)/2 and Py± = (I ± Y )/2 are projectors onto eigenstates of X and
P j j
Y respectively. Therefore each term in eiµ0 j Ol Or is of the form of
P a a
a Pl Ur , where
each measurement is applied only on a single qubit. In standard teleportation, one usually
implements a Bell pair measurement rather than a single qubit measurement in the com-
putational basis. The extra simplicity in our situation is possible because of the scrambling
generated by time evolution [9].
If we include Majorana fields with both 2j − 1 and 2j indices in the coupling V ,
their projections will not commute, as shown in (2.9) and (2.10). Therefore, to relate to
teleportation with a classical channel, one should consider coupling at most half of the N
Majoranas, as discussed in Sec. 5.2.
After the three steps, the state becomes

ST r USQl |φ0 iRQ |tfdilr |0iT

1   
= √ |0iT ⊗ χr χ†r + |1iT ⊗ χr U |00iRQ ⊗ χl χ†l
2

+ |01iRQ ⊗ χl + |10iRQ ⊗ χ†l + |11iRQ ⊗ χ†l χl |tfdilr (2.11)

where χr , χ†r are short for χr (t), χ†r (t) and χl , χ†l are short for χl (−t0 ), χ†l (−t0 ). In the main
part of this paper, we suppress their time arguments for notational simplicity.
In order to quantify the success of teleportation, we compute the mutual information
between R and T. Perfect teleportation corresponds to

IRT = S(R) + S(T ) − S(RT ) = 2 log 2 (2.12)

The reduced density matrix of TR is

 
ρ11 0 0 ρ14
1 0
 ρ22 ρ23 0 
ρT R =  (2.13)

2 0 ρ∗23

ρ33 0 
ρ∗14 0 0 ρ44
where the trivial matrix elements are due to fermion number conservation
D E
χi1 · · · χis χ†is+1 · · · χ†i2k+1 = 0, k ∈ N (2.14)

and the nontrivial matrix elements are given as

D E D E
ρ11 = χl χ†l U † χr χ†r Uχl χ†l + χ†l U † χr χ†r Uχl (2.15)
D E D E
ρ14 = χl U † χ†r Uχl χ†l + χ†l χl U † χ†r Uχl (2.16)
D E D E
ρ22 = χl U † χr χ†r Uχ†l + χ†l χl U † χr χ†r Uχ†l χl (2.17)
D E D E
ρ23 = χl χ†l U † χ†r Uχ†l + χ†l U † χ†r Uχ†l χl (2.18)
D E D E
ρ33 = χl χ†l U † χ†r χr Uχl χ†l + χ†l U † χ†r χr Uχl (2.19)
D E D E
ρ44 = χl U † χ†r χr Uχ†l + χ†l χl U † χ†r χr Uχ†l χl (2.20)

–7–
where U ≡ exp(iµV ) and all expectation values are taken in thermofield double state. A
consistency check is that if µ = 0, then we get ρT R = 12 diag(ρ11 , ρ11 , 1 − ρ11 , 1 − ρ11 ) which
leads to zero mutual information IRT .
Perfect teleportation (2.12) requires that both S(R) = S(T ) = 2 log 2 and S(RT ) = 0.
Given the structure of ρT R in (2.13), the first conditions imply that

ρ11 = ρ44 , ρ11 + ρ22 = ρ11 + ρ33 = 1 (2.21)

and second condition implies that either

(
ρ11 = |ρ14 | = 1, |ρ23 | = 0
(2.22)
ρ11 = |ρ14 | = 0, |ρ23 | = 1
Consider the first case. It is obvious that
D E D E D E
ρ11 + ρ33 = χl χ†l U † {χr , χ†r }Uχl χ†l + χ†l U † {χr , χ†r }Uχl = {χl , χ†l } = 1 (2.23)

and similarly ρ22 + ρ44 = 1. Therefore, ρ11 = 1 is equivalent to ρ33 = 0, which implies that
the two terms in (2.19) vanish, as both are nonnegative

χr Uχl χ†l |tfdi = χr Uχl |tfdi = 0 (2.24)

Similarly, ρ22 = 0 implies that

χ†r Uχ†l |tfdi = χ†r Uχ†l χl |tfdi = 0 (2.25)

Based on these, ρ23 is automatically trivial. The only remaining nontrivial condition is
|ρ14 | = 1. From (2.24) and (2.25), we can derive

χr Uχ†l χl |tfdi = χr U |tfdi , (2.26)

χ†r Uχl χ†l |tfdi = χ†r U |tfdi (2.27)

which can be used to simplify |ρ14 | = 1 to the condition

D E
 {U † χ†r U, χl }  = 1 (2.28)
 

This has a clear physical interpretation: the left-right causal propagator reaches its max-
imum value. Note that because we are discussing fermions, the propagator can never be
larger than 1.
For the second possible case above, the condition for perfect teleportation is similar

ρ11 = 0 =⇒ χ†r Uχl χ†l |tfdi = χ†r Uχl |tfdi = 0 (2.29)

ρ44 = 0 =⇒ χr Uχ†l |tfdi
= χr Uχ†l χl |tfdi = 0 (2.30)
D E
|ρ23 | = 1 =⇒  {U † χr U, χl }  = 1 (2.31)
 

Comparing the two, one can see that the difference is just an exchange of χr and χ†r . This
just corresponds to flipping the convention of which state on the right we call 0 versus 1.
If we interpret χ† as creation of a particle and χ as annihilation of a particle, it is natural
to take the first case.

–8–
2.2 SYK model
In the SYK model, we can construct Dirac fermions from two Majorana fermions. More
precisely, we will consider an even number N of Majorana fermions on each side:
j
Jjl,r ψ j1 · · · ψl,rq ,
X
Hl,r = iq/2 1 ···jq l,r
(2.32)
1≤j1 <···<jq ≤N

2q−1 J 2 (q − 1)! J 2 (q − 1)!

 2 
Jjl,r
1 ···jq
= = (2.33)
qN q−1 N q−1

where the normalization is

j k
{ψl,r , ψl,r } = δ jk (2.34)
Let |Ii be the maximally entangled state of l and r defined by

(ψli + iψri ) |Ii = 0, ∀i (2.35)

In other words, |Ii is annihilated by a complex fermion fj = √1 (ψ j + iψrj ).

2 l
The thermofield double state is then given by
−1/2 −βH/4
|tfdi = Zβ e |Ii , H ≡ Hl + Hr (2.36)

We further require that Hl − Hr |Ii = 0 so this state is the infinite temperature entangled
state. Using (Hl − Hr ) |Ii = 0, this implies that the Hamiltonia must be related by

Jjl1 ···jq = iq Jjr1 ···jq (2.37)

The SYK model can be solved using path integrals after averaging over the Gaussian
ensemble of random coupling J. Note that the thermofield double state is defined by a
Euclidean evolution of length β/4 from |Ii. Euclidean time ordered correlation functions
in the thermofield double state are related to
D E
I|e−βH/2 TO1 · · · Ok |I (2.38)

which is equivalent to the path integral over a length of β/2.4 Therefore, the action we
start with is
Z β/2 " X N
#
1 i i i i
S= dτ (ψl ∂τ ψl + ψr ∂τ ψr ) + Hl + Hr (2.39)
0 2
i=1

Since we have two copies with random couplings related by (2.37), integrating over J will
give interaction terms between ψl and ψr [16]. Introducing four delta functions
N
!
1 X i i
δ Gab (τ1 , τ2 ) − ψa (τ1 )ψb (τ2 ) (2.40)
N
i=1
4
If we take all Oi at β/4, this is the expectation value in the thermofield double state.

–9–
for a, b = l, r in the path integral, we find
Z β/2 N Z β/2 N
1 X 1X X
S= dτ (ψli ∂τ ψli + ψri ∂τ ψri ) − dτ dτ 0 Σab (τ, τ 0 ) ψai (τ )ψbi (τ 0 )
2 0 2 0
i=1 a,b i=1
β/2
J2
XZ  
N
+ dτ dτ 0 Σab (τ, τ 0 )Gab (τ, τ 0 ) − sab Gab (τ, τ 0 )q (2.41)
2 0 q
a,b

where sll = srr = 1 and slr = srl = (−)q/2 . We stitch together the left and right fermions
into a single field, defining ψ for τ ∈ [0, β] as
( j
ψl (τ ) τ ∈ [0, β/2]
ψ j (τ ) = j
(2.42)
iψr (β − τ ) τ ∈ [β/2, β]

and corresponding G, Σ fields

G(τ, τ 0 ), Σ(τ, τ 0 )



Gll (τ, τ 0 ), Σll (τ, τ 0 ) τ, τ 0 ∈ [0, β/2]

iG (τ, τ 0 ), −iΣ (τ, τ 0 )

τ, τβ0 ∈ [0, β/2]
lr β lr β
= (2.43)


iGrl (τβ , τ ), −iΣrl (τβ , τ 0 )
0 τβ , τ 0 ∈ [0, β/2]

−G (τ , τ 0 ), −Σ (τ , τ 0 )

τβ , τβ0 ∈ [0, β/2]
rr β β rr β β

where τβ ≡ β − τ . The action then simplifies to

N Z β Z β 
1X i i 0 i 0 i 0
S= dτ ψ ∂τ ψ − dτ dτ ψ (τ )Σ(τ, τ )ψ (τ )
2 0 0
i=1
β
J2
Z  
N 0 0 0 0 q
+ dτ dτ Σ(τ, τ )G(τ, τ ) − G(τ, τ ) (2.44)
2 0 q
which unsurprisingly becomes the path integral of a single field ψ on the thermal circle of
length β. The only thing to check in this treatment is that on the stitching boundary β/2,
j
the field redefinition may not be smooth. Fortunately, if we extend the definition of ψl,r
beyond τ = β/2 via Euclidean evolution, we have

hI| e−βH/2 ψlj (β/2 + τ ) = hI| eHτ ψlj e−(β/2+τ )H

= hI| ψlj e2Hr τ e−(β/2+τ )H = i hI| e−βH/2 ψrj (β/2 − τ ) (2.45)

thanks to (2.35). This shows that the fields in the path integral are smoothly defined at
τ = β/2. Moreover, one can similarly show that

ψlj (−τ ) |Ii = −iψrj (τ ) |Ii (2.46)

which means that we can even extend the path integral past τ = β by requiring the usual
thermal anti-periodic ψ i (τ ) = −ψ i (β + τ ) boundary condition. Integration over ψ i results
in an effective action for the G and Σ fields.
ZZ
N N 1
S[G, Σ] = − log det(∂τ − Σ) + (ΣG − J 2 Gq ) (2.47)
2 2 q

– 10 –
In the large N limit, the equation of motion is

G−1 = G−1 0 2 0 q−1

0 −Σ, Σ(τ, τ ) = J G(τ, τ ) , (2.48)
1
G−1
0 = ∂τ ⇐⇒ G0 (τ ) = sgnτ (2.49)
2
We will add instantaneous interactions between the two SYK systems at time τ0 by
[16]
N N
i X j j µ X † µ
Hint = −µV = −µ × ψl ψr = − fj fj + (2.50)
qN qN 2q
j=1 j=1

We will take µ to be O(1) in the 1/N counting, as in previous work [2, 26]. Thus µV has
an O(1) effect on the state |Ii, namely
µ
µV |Ii = − |Ii (2.51)
2q
Writing Hint in terms of G field, the action is shifted by
µ
δS = − [G(τ0 , β − τ0 ) − G(β − τ0 , τ0 )] (2.52)
2q
Such an instantaneous interaction does not affect the equation of motion but only changes
the boundary condition of G and Σ at τ0 (see Sec. 3.1.2).
It turns out that
µ 
Σ(τ, τ 0 ) =J 2 G(τ, τ 0 )q−1 +

δτ,τ0 δτ 0 ,β−τ0 − δτ,β−τ0 δτ 0 ,τ0 (2.53)
qN

where δa,b is short for δ(a − b). On the other hand, G−1 = G−1 0 − Σ can be written as
Z
∂τ 0 G(τ 0 , τ ) − dτ 00 Σ(τ 0 , τ 00 )G(τ 00 , τ ) = δ(τ − τ 0 ) (2.54)

Integrating τ 0 around τ0 and β − τ0 respectively, we get

µ
G(τ0+ , τ )−G(τ0− , τ ) = G(β − τ0 , τ ) (2.55)
qN
µ
G((β − τ0 )− , τ )−G((β − τ0 )+ , τ ) = G(τ0 , τ ) (2.56)
qN
where subscript in τ± means τ ± ε for ε → 0. Note that the RHS of above equations are not
defined a priori, since τ0 and β − τ0 are singular points. Instead these two equations should
be taken as the definition of the values of G at these two points, as follows. Consider the
expansion of G as a series in powers of µ, so at leading order G is continuous. Hence, the
discontinuity of G around τ0 and β − τ0 at linear order in µ is given by the zero-th order
value of G on the RHS.5 This is exactly the linearized version of twist boundary condition
of G we will derive from the microscopic fermion description in Sec. 3.1.2.
5
Beyond linear order, it is not clear how to determine the discontinuity directly from (2.55) and (2.56).

– 11 –
2.3 Density matrix ρT R in SYK model
For each of the pair of SYK systems, we construct Dirac fermions by combining two Ma-
jorana fermions
1 1
χjl,r = √ (ψl,r
2j−1 2j
+ iψl,r ), χj† 2j−1
l,r = √ (ψl,r
2j
− iψl,r ) (2.57)
2 2
with commutation relations
{χil,r , χj†
l,r } = δ
ij
(2.58)
It follows that
1 1
χjl,r χj† − iψl,r
l,r =
2j−1 2j
ψl,r , χj† j 2j−1 2j
l,r χl,r = 2 + iψl,r ψl,r (2.59)
2
Using above constructions, we can write down the matrix element of ρT R more explic-
itly. In this subsection, we take χ1l,r as the Dirac fermion appearing in the general protocol
described in section 2.1. Take ρ11 as an example. The first term is
D E 1D E iD E 1D E
χ1l χ1† † 1 1† 1 1†
l U χr χr Uχl χl = χ1l χ1† l − U † ψr1 ψr2 U − {ψl1 ψl2 , U † ψr1 ψr2 U}
2 D 4 E 2
+ i ψl1 ψl2 U † ψr1 ψr2 Uψl1 ψl2 (2.60)
The second term is
D E 1D E iD E 1D E
χ1† † 1 1† 1 1† 1 1 † 1 2 1
l U χr χr Uχl = 2 χl χl − 2 ψl U ψr ψr Uψl − 2 ψl U ψr ψr Uψl
2 † 1 2 1

1 D 1 † 1 2 2E i D 2 † 1 2 2E
+ ψl U ψr ψr Uψl − ψl U ψr ψr Uψl (2.61)
2 2
For the SYK model, after performing the ensemble average, there is an emergent
SO(N ) symmetry that rotates all ψl,ri by U ij ψ j . This symmetry must be manifested by
l,r
correlation functions, thus

i1
ψ · · · ψ i2k ∝ δ π1 π2 · · · δ π2k−1 π2k

(2.62)
where πi is the permutation of {i1 , · · · , i2k }. As V contains all pairs of ψli and ψri with the
same indices, the correlation functions in (2.60) and (2.61) with an odd number of 1 and 2
indices vanish. Moreover, this symmetry dictates that correlation functions are invariant
under exchange of the 1 and 2 indices. Therefore, we get
1h D Ei
ρ11 = 1 − {ψl1 , [ψl2 , U † ψr1 ψr2 U]} (2.63)
2
Similarly, we can work out all other matrix elements of ρT R . They are
1D 1 † 1 E D 2 1 † 2 1 E
ρ14 = {ψl , U ψr U} − {ψl , ψl U ψr Uψl } (2.64)
2
1 h D Ei
ρ22 = 1 + {ψl1 , [ψl2 , U † ψr1 ψr2 U]} = 1 − ρ11 (2.65)
2
ρ23 = 0, ρ33 = 1 − ρ11 , ρ44 = 1 − ρ22 (2.66)
Note that in (2.66) ρ23 becomes trivial exactly due to the SO(N ) symmetry, while ρ33
and ρ44 are related to ρ11 and ρ22 by definition. Comparing with the conditions for perfect
teleportation (2.21) and (2.22), the only nontrivial conditions for obtaining maximal mutual
information between R and T are
ρ11 = |ρ14 | = 1 (2.67)

– 12 –
3 Computing the correlation functions in the SYK model

In this section, we will use the method developed in [27]. To calculate ρ11 and ρ14 , there
are two types of correlation functions that need to be found

e−iµV ψ 1 eiµV ψ 1 , e−iµV ψ 1 ψ 2 eiµV ψ 1 ψ 2



(3.1)

where we will consider all orderings of the operators. We first focus on the former. For
notational simplicity, in this section we use ψ to represent ψ 1 .

3.1 Stitching correlation function

With the stitching field ψ defined in (2.42), we define a stitching correlation function for
τa , τb ∈ [0, β] as

−βH/2  −µV (β/4+) µV (β/4−)  
I|e T̄ e e ψ(τa )ψ(τb ) |I
Gµ (τa , τb ) =
 sgn(τ̄a − τ̄b ) (3.2)
I|e−βH/2 e−µV (β/4+) eµV (β/4−) |I

where τ̄ ≡ min(τ, β − τ ) and T̄ represents time ordering for τ̄a and τ̄b (without any ± sign
from (anti-)commutation of ψ). Here  ∈ (0, β/4) is a regulator that we will ultimately
take to zero. As discussed in Sec. 2.2, this correlation function is smoothly defined around
β/2. Note that Gµ clearly can be extended to any real τa and τb using the periodicity of
ψ, ψ(τ + β) = −ψ(τ ). We will call τa , τb ∈ [0, β] the basic domain (see Fig. 2).

3.1.1 Symmetries
There are several symmetries of Gµ . It is anti-symmetric in τa and τb

Gµ (τa , τb ) = −Gµ (τb , τa ) (3.3)

which together with the periodicity can be summarized as

Gµ (τa , τb ) = −Gµ (τb , τa ) = (−)n+m Gµ (τa + nβ, τb + mβ) (3.4)

Another feature is that Gµ is real. This can be seen as follows. First, the Hamiltonia
Hl,r have only real coefficients and µ is real. Second, ψlj = √12 (fj +fj† ) and iψrj = √12 (fj −fj† )
are linear combinations with real coefficients. Hence we can easily conclude that as op-
erators both e−βH/2 T̄ e−µV (β/4+) eµV (β/4−) ψ(τa )ψ(τb ) and e−βH/2 e−µV (β/4+) eµV (β/4−)
 

consist of ci and c†i with real coefficients. Thus Gµ must be real.

This leads to further symmetries. Define Ol = ψl , Or = iψr and
h i
C± AB
(τ1 , τ2 ) = T e−µV (β/4±) eµV (β/4∓) OA (τ1 )OB (τ2 ) (3.5)

For τ̄a , τ̄b ∈ [0, β/2] and A, B = l, r, we have

D E∗ D E
I|e−βH/2 C+AB
(τ̄a , τ̄b )|I = (−1)s I|e−βH/2 C−
AB
(β/2 − τ̄a , β/2 − τ̄b )|I (3.6)

– 13 –
where s = 0 if A = B and s = 1 if A 6= B (because there is an i with ψr ), and T is ordinary
time ordering. Here we used the fact
†
O(τ )† = eHτ Oe−Hτ = e−Hτ OeHτ = O(−τ )

(3.7)

This implies that in the basic domain τa , τb ∈ [0, β]

Gµ (τa , τb ) = G−µ (β/2 − τb , β/2 − τa ) (3.8)

Using this symmetry and shifts by β, we get another symmetry

Gµ (τa , τb ) = G−µ (3β/2 − τa , β/2 − τb ) (3.9)

= G−µ (3β/2 − τb , 3β/2 − τa ) (3.10)

Finally, if q ∈ 4Z there is one more discrete symmetry of Gµ resulting from

ψlj → ψrj , ψrj → −ψlj (3.11)

This can be seen by noting that Gµ is the expectation of a complicated operator that can
be expanded as a sum of various products of ψlj and ψrk in the state |Ii. Since |Ii is defined
by (ψlj + iψrj ) |Ii = 0 for all j, |Ii has a tensor product form ⊗j |Ij i. Therefore, we only
need to check if all correlations in each |Ij i are invariant under (3.11). It is clear that for
all j
D E i D E

Ij |ψlj ψrj |Ij = , Ij |ψlj |Ij = Ij |ψrj |Ij = 0


(3.12)
2
and all higher point correlations can be reduced to them using (2.34). This means that
whatever Gµ is, the whole expression is invariant under (3.11). Second, V is invariant
X j
ψrj (−ψlj ) = i
X X j
V =i ψl ψrj = i ψl ψrj (3.13)
j j j

Third the Hamiltonian H is invariant when q ∈ 4Z as then J l = J r . Hence we have the

following identity

Gµ (τa , τb ) = −Gµ (β − τa , β − τb ) = Gµ (β − τb , β − τa ) (3.14)

These symmetries kill some regions in the left top triangle in the basic domain. Indeed,
(3.8) and (3.10) says τa + τb = β/2, 3β/2 are two invariant lines, which kill the regions
A and B in Fig. 2. (3.9) implies the existence of a reflection symmetry around point
(τa , τb ) = (3β/4, β/4) which kills region C. (3.14) says τa + τb = β is an invariant line,
which kills region D. In the end, the only independent region is the dark yellow square in
Fig. 2, which we will call the fundamental region.

– 14 –
Figure 2: The stitching correlation function Gµ in the basic domain is split into a few
small regions. Due to symmetries, the independent region is the dark yellow one, which
we call the fundamental region. Blue dashed lines are invariant under symmetries of Gµ ,
which implies that the white dotted regions A, B, C, D are not independent. The twisted
boundary condition is on the four red lines, which split the fundamental domain further
into five subregions a, b, c, d, e.

3.1.2 Twist boundary condition

The boundary conditions on the fermions induced by the instantaneous interaction result
from the following identity
! ! !
µ µ
µV ψl −µV cosh qN − sinh qN ψl
e e = µ µ (3.15)
iψr − sinh qN cosh qN iψr
When ψ moves across β/4 ± , it gets acted upon by the above connection equation (and
its −µ version). It follows that for the stitching field ψ
! ! !
µ µ
ψ(τ ) cosh qN ∓ sinh qN ψ(τ )
lim = µ µ lim (3.16)
τ →(β/4∓)− ψ(β − τ ) ∓ sinh qN cosh qN τ →(β/4∓)+ ψ(β − τ )

As Gµ contains two ψ insertions, this gives a twist boundary condition

  ! 
lim Gµ (τa , τb ) µ µ lim Gµ (τa , τb )
cosh qN ± sinh qN
 τa ,τb →(β/4∓)+ = µ µ
 τa ,τb →(β/4∓)− 
lim Gµ (τa , τb ) ± sinh qN cosh qN lim Gµ (τa , τb )
τa ,τb →(3β/4±)− τa ,τb →(3β/4±)+
(3.17)
Using the symmetries we can rewrite everything in terms of twist boundary conditions in
fundamental region. Indeed, we have
Gµ (τa , (3β/4 ± )− ) = −Gµ (β − τa , (β/4 ∓ )+ ) (3.18)
Gµ ((β/4 − )± , τb ) = −G−µ ((β/4 + )∓ , β/2 − τb ) (3.19)
Gµ ((3β/4 + )∓ , τb ) = G−µ ((3β/4 − )± , β/2 − τb ) (3.20)

– 15 –
Figure 3: The smoothness condition for Gµ . Gµ needs to be smooth in each region bounded
by dashed and solid red lines. On dashed red lines, Gµ must satisfy a twist boundary
condition, and on solid red lines, Gµ need not be smooth. Blue dashed lines are symmetry
lines on which Gµ must be smooth. Due to the symmetry, we only need to solve for Gµ on
the five regions a, b, c, d, e.

It follows that the independent twist boundary conditions in fundamental region are
! ! !
µ µ
Gµ ((β/4 + )+ , τb ) cosh qN − sinh qN Gµ ((β/4 + )− , τb )
= µ µ (3.21)
Gµ ((3β/4 − )− , τb ) − sinh qN cosh qN Gµ ((3β/4 − )+ , τb )

and
µ µ
Gµ (τa ,(β/4 ∓ )+ ) = cosh Gµ (τa , (β/4 ∓ )− ) ∓ sinh Gµ (β − τa , (β/4 ∓ )− ) (3.22)
qN qN
They are on the red dashed lines in Fig. 2, and separate the fundamental region into
five subregions a, b, c, d, e. One can check that at linear order in 1/N , the twist boundary
conditions are the same as (2.55) and (2.56).

3.1.3 Smoothness and UV conditions

Although the symmetries reduced the problem to the fundamental region, that does not
mean that we can take an arbitrary function that satisfies the twist boundary conditions
and simply transform it to the other regions via the symmetry. In particular, the correlation
function Gµ should be further constrained by smoothness and UV conditions.
The smoothness condition is that Gµ must be a smooth function of τa,b when no op-
erators coincide with each other. It follows that Gµ must be smooth crossing certain
symmetry lines, for example, the lines τa = ±τb + β/2. On the other hand, some symmetry
lines are “hard” lines that correspond to coincident operators, for example, τa = ±τb and
τa = ±τb + β. In the latter case, Gµ need not be smooth and the extension across them is
simply to “copy and paste” by the symmetry.

– 16 –
 In doing this, we need to study Gµ on a bigger patch as shown in Fig. 3. In this figure,
Gµ needs to be smooth in each region bounded by dashed and solid red lines. On dashed
red lines, Gµ must satisfy the twist boundary condition, and on solid red lines, Gµ need not
be smooth. Blue dashed lines are symmetry lines on which Gµ must be smooth.
The input from the UV is that when τa → τb , Gµ must equal 1/2, which are required
by the Majorana fermion algebra. This gives a boundary condition on the line τa = τb .
Due to the symmetry, we only need to solve for Gµ on the new five subregions a, b, c, d, e
in Fig. 3.

3.2 Large q solution

3.2.1 General solution
The insertion of e±µV in the correlation function Gµ can be understood as adding a defor-
mation to action at β/4 ±  of the form of (2.52). This does not change the bulk equation
of motion (2.48), but simply imposes the twist boundary condition we analyzed above.
If we fix J , the large q limit will truncate (2.48) to first order [15, 28, 29] as
J2
[G] = [G0 ] + [G0 ][Σ][G0 ], Σ= (2G)q−1 (3.23)
q
where [A][B] means product by integrating over the intermediate variable τ . We make the
ansatz for τab > 0
J 2 σ(τa ,τb )
G(τa , τb ) = G0 (τab )eσ(τa ,τb )/(q−1) =⇒ Σ(τa , τb ) = e (3.24)
q
Taking time derivative of [G] in (3.23), we get
J 2 σ(τa ,τb )
∂a ∂b (G − G0 ) = − e , τab > 0 (3.25)
q
In the large q limit, this becomes
∂a ∂b σ = −2J 2 eσ(τa ,τb ) (3.26)
which is the well-known Liouville equation. The most general solution is
f 0 (τa )g 0 (τb )
σ = ln (3.27)
(1 + J 2 f (τa )g(τb ))2
It follows that the general solution for G in the large q limit is
 1
f 0 (τa )g 0 (τb )

q−1
G(τa , τb ) = G0 (τab ) (3.28)
(1 + J 2 f (τa )g(τb ))2
This solution has a global SL(2) symmetry. Under the following transformation
a + bf (x) d − cJ 2 g(x)
f (x) → , g(x) → 2 , (3.29)
c + df (x) J (−b + aJ 2 g(x))
where bc − ad = 1, (3.27) is invariant.6 This global SL(2) symmetry is a gauge redundancy
in the expression of G. In the following, we will discuss the solution for G modulo this
redundancy.
6
Note that overall scaling of a, b, c, d does not change f , and thus we only the SL(2) subgroup of GL(2)
acts non-trivially.

– 17 –
3.2.2 Translation invariant solution: µ = 0 case
When µ = 0, Gµ = G becomes the usual time translation invariant thermal correlation
function. This means that G(τa , τb ) is a function of τ = τab only. In particular, ∂b = −∂a
in the differential equation (3.26) and we get

∂τ2 σ(τ ) = 2J 2 eσ(τ ) (3.30)

The general solution to this is

ω2
eσ(τ ) = (3.31)
J 2 cos2 (ωτ + v)

where ω and v are two integral constants. If we set τb = 0 and τa = τ , we can compare it
with (3.27) and find
A + B tan(ωτ + v)
f (τ ) = (3.32)
C + D tan(ωτ + v)
where A, B, C, D are some constants involving g(0). We fix the SL(2) symmetry by choos-
ing
1
f (τ ) = tan(ωτ + v1 ) (3.33)
J
1
Assume g(τ ) = J tan ωg̃(τ ). Plugging this into (3.27) we find

ω 2 g̃ 0 (τb )
eσ(τab ) = (3.34)
J 2 cos2 (ωτa − ωg̃(τb ) + v1 )

Comparing with (3.31), it is easy to see that g̃(τb ) = τb + v2 for some constant v2 .
In the following, Sec. 3.2.3, and Appendix A, we will shift τa to τa → τa + β/2
to simplify the action of the symmetries. In these shifted coordinates, the symmetries
discussed in subsection 3.1.1 now become (as µ = 0)

(f (τa ), g(τb )) ' (g(τa ), f (τb )) ' (g(−τa ), f (−τb )) (3.35)

where ' means equivalent under global SL(2) transformation (3.29), namely the σ value
is the same. This restricts v2 = v1 = v. We summarize the translation invariant solution
as follows
tan(ωτ + v) σ ω2
f (τ ) = g(τ ) = , e = 2 (3.36)
J J cos2 ωτab

The UV condition states that when τa = τb − β/2, we must have that eσ = 1. This
determines
ω = J cos ωβ/2 (3.37)

– 18 –
3.2.3 Twist solution at large q
Let us move to nonzero µ case. The twist boundary condition makes it hard to find f and
g across the various subregions of the fundamental region. However, as the coupling scales
as 1/q, the twist boundary condition significantly simplifies in the large q limit. At order
1/q, (3.22) and (3.21) become

σ(τa , (β/4 ∓ )+ ) = σ(τa , (β/4 ∓ )− ) ∓ µ/N (3.38)

σ((−β/4 + )+ , τb ) = σ((−β/4 + )− , τb ) − µ/N (3.39)
σ((β/4 − )− , τb ) = σ((β/4 − )+ , τb ) − µ/N (3.40)

which means that two twisted components decouple. All of these conditions can be written
in the following form

f10 (τa )g10 (τb ) e±µ/N f20 (τa )g20 (τb )

= (3.41)
(1 + J 2 f1 (τa )g1 (τb ))2 (1 + J 2 f2 (τa )g2 (τb ))2
where one of τa,b is fixed to be the value on a border between subregions, and subscripts
1 and 2 are used to distinguish the functions f and g on neighboring subregions (for
example, 1 could refer to subregion d, 2 could refer to subregion c, and we choose minus
sign in (3.41)). Let us fix τb as in (3.41), and then we can integrate τa to get

g10 (τb ) e±µ/N g20 (τb )

= +c (3.42)
J 2 g1 (τb )(1 + J 2 f1 (τa )g1 (τb )) J 2 g2 (τb )(1 + J 2 f2 (τa )g2 (τb ))

One can easily find the relation between f1 (τa ) and f2 (τb ) across the border: they are
related to each other via an SL(2) transformation that depends on the values of g1,2 on
the border. Note that this boundary condition does not give any information about how
g1 (τb ) is related to g2 (τb ). Since we have a global SL(2) symmetry, we can do a SL(2)
transformation for f2 and g2 such that f1 = f2 . A similar analysis applies to other borders
to fix b, c and d with same f µ and a, c and e with same g1µ (see Fig. 4).
Note that now the symmetry lines are τa = ±τb , across which we have

Gµ (τa , τb ) = G−µ (τb , τa ), Gµ (τa , τb ) = G−µ (−τb , −τa ) (3.43)

Together with the smoothness requirement, this strongly constrains the function choice in
subregion b (see Fig. 4). Suppose the function in b is denoted as (f0µ , g0µ ). Then (3.43)
implies that
(f0µ , g0µ ) ' (g0−µ , f0−µ ), (f0µ , g0µ ) ' (g0∗
−µ −µ
, f0∗ ) (3.44)
where ∗ denotes the flip τ → −τ . It follows that g0µ must be a SL(2) transform of f0−µ .
Let us assume g0µ = SLf µ (f −µ ) which obeys (f µ , SL
f µ (f −µ )) ' (SL
f −µ (f µ ), f −µ ) due to the
0 0 0 0 0
first symmetry in (3.43). When µ = 0, we see from (3.36) that the gluing SL f reduces to
the identity so that the solution is smooth.
Similarly, f0µ must be related to f0∗
µ
by an SL(2) transformation, as f0µ = slµ (f0∗ µ
), due
to the second symmetry in (3.43) and this SL(2) transformation should also be compatible
with SLf µ . This imposes a few constraints on both SL(2) transformations. Such an analysis

– 19 –
Figure 4: The assignment of different (fi , gi ) functions to each region with all symmetry
conditions satisfied in the large q case. We need to determine three functions f µ , g1µ and
g µ with twist boundary condition on the blue rectangle, and four SL(2) transformations
SL
f µ , SL
c µ , SLµ and slµ .

can be applied to regions a and e, in which each have only one symmetry. Taking all these
symmetries in to account, we only need to determine three functions f µ , g1µ and g µ with
twist boundary condition on the blue rectangle in Fig. 4, and four SL(2) transformations
SL
f µ , SL
c µ , SLµ and slµ .
The twist boundary conditions on the four blue segments in Fig. 4 are given as follows

g1µ 0 (τ+ ) = e−µ g µ0 (τ+ ), g1µ (τ+ ) = g µ (τ+ ), (3.45)

f µ (f −µ )0 (τ− ) =
SL eµ g1µ 0 (τ− ), SL
f µ (f −µ )(τ− ) = g µ (τ− ),
1 (3.46)
c µ (g −µ )0 (τ− )
SL = c −µ
e f (τ− ), SLµ (g1 )(τ− ) = f µ (τ− ),
µ µ0
(3.47)
1
−µ 0 −µ
SLµ (g1∗ ) (−τ− ) = eµ f µ0 (−τ− ), SLµ (g1∗ )(−τ− ) = f µ (−τ− ) (3.48)

where τ± ≡ β/4±. We can use these equations to show that SL c µ is completely determined
by SL
f µ . See Appendix A for more details.
Let us discuss the UV constraint, which requires that eσ = 1 when τa = τb − β/2.
In general, this condition is surprisingly easy to satisfy. In the unshifted coordinate τa , it
means that
f 0 (τ )g 0 (τ )
=1 (3.49)
(1 + J 2 f (τ )g(τ ))2
Indeed, for arbitrary f and g, we need only to choose τ as the proper length in the space
(f, g) ∈ R2 with metric
df dg
dτ 2 = (3.50)
(1 + J 2 f (τ )g(τ ))2

– 20 –
Therefore, in principle we may expect infinitely many solutions for (f µ , g1µ , g µ ) satisfying
all of the conditions of our problem. Indeed, this is because the large q equation of motion
(3.23) is too loose. The constructive approach would require working to the next order in
1/q, which should lift the flat direction of solutions.
However, to get find any solution analytically is not easy. For example, it is not clear
that after redefining τ according to UV requirement, the functions in region e will be still
−µ
of the form (SLµ (g1∗ ), g1µ ), namely that the first component is related to the second one
via a reflection and SL(2). In order to find a reasonably simple solution, we make further
assumptions:
Aµ + Bµ f f
f µ = f, g1µ = G(f ) ≡ , SLµ = I (3.51)
Cµ + D µ f
which also implies that SL c µ = I. The first assumption f µ = f is quite strong, but it is
a good candidate to satisfy the strong symmetry constraints in subregion b. The second
assumption is inspired by (3.15) where moving across eµV is a SL(2) transformation for ψ.
An important consistency criterion on the solution is that when it is continued to
Lorentz signature and the limit  → 0 is taken, it should obey causality. In other words,
the correlation functions in the past of the interaction eigV must be the original ones of the
thermofield double state in the decoupled system. We will see later that our solution has
this property. For example in the region b, with our ansatz ψl (t0 )eigV e−igV ψr (t) is the



original translation invariant two point function. However in region d, only after taking
 → 0 can one see that the Lorentzian solution obeys causality.
With this ansatz, solving for (fi , gi ) in all subregions is straightforward. We leave the
details to Appendix A and only describe the result here. Taking τ+ = τ− = β/4 and
shifting τa back τa → τa − β/2, the solutions of eσ in each subregion are as follows:
"
e−2µ/N ω 2 β (e−µ/N − 1)
[eσ ](a) = cos ω(τab − ) + cos ωτab
J2 2 cos ωβ/2
#−2
(e−µ/N − 1)2 sin ωβ/2 β β
+ cos ω(τa − ) sin ω(τb − ) (3.52)
cos2 ωβ/2 4 4
ω2
[eσ ](b) = [eσ ](e) = [eσ ](d) = (3.53)
J 2 cos2 ω(τab − β2 )
e−µ/N ω 2
[eσ ](c) =  2 (3.54)
β (e−µ/N −1) β β
J2 cos ω(τab − 2) + cos ωβ/2 sin ω(τa − 4 ) sin ω(τb − 4)

It is interesting that the solution in subregion c has the same form as that in [27].

4 Effectiveness of protocol

In order to check the mutual information IRT after performing the protocol, we need to
evaluate correlation functions in Lorentz signature. The Lorentzian correlation functions
are analytic continuations of Gµ . For example, the OTOC U † ψl1 (t)Uψl1 (t) is equal to



lim Giµ (β/4 + it, (β/4 − )− + it) (4.1)

→0

– 21 –
4.1 Lorentzian correlation function
To find the density matrix ρij , we need to calculate a few OTOCs. We first consider
D E D E D E
K(t, t0 ) = {ψl (−t0 ), U † ψr (t)U} = ψl (−t0 )U † ψr (t)U + U † ψr (t)Uψl (−t0 ) (4.2)

Comparing with the definition of Gµ , we can easily find that

K(t, t0 ) = −i lim Giµ (−it0 + β(1/4 + )+ , 3β/4 − it) + Giµ (3β/4 − it, −it0 + β(1/4 − )− )
 
→0
 
i e−iµ/(N q) ω 2/q
= i2/q − h.c. (4.3)

2
 h
J (e−iµ/N −1)
J 2/q cosh ω∆t − ω sinh ω(t − iβ/2) sinh ωt0

where we used symmetries from Sec. 3.1.1 to express Gµ in the fundamental domain and
in the last line ∆t ≡ t0 − t.
There is another type of OTOC involved in ρij , namely ψ 1 ψ 2 U † ψ 1 ψ 2 U . One can



follow a similar procedure as in the calculation of G to derive the resulting twist boundary
condition for four point correlation functions

−βH/2  −µV (β/4+) µV (β/4−) 1
ψ (τa )ψ 1 (τb )ψ 2 (τc )ψ 2 (τd ) |I
 
I|e T̄ e e
Ĝ(τa , τb ; τc , τd ) =
 sgn(τ̄a , τ̄b , τ̄c , τ̄d )
I|e−βH/2 e−µV (β/4+) eµV (β/4−) |I
(4.4)
where sgn(τ̄a , τ̄b , τ̄c , τ̄d ) = 1 for τ̄a > τ̄b > τ̄c > τ̄d and the sign flips for every permutation
of the order. It is very straightforward to see that for fixed τc and τd , the correlator as a
function of τa and τb obeys the same twist boundary condition when crossing e±µV as before.
The same applies to the boundary condition for τc and τd with τa and τb fixed. Therefore,
we can understand Ĝ as a four point function I|e−βH/2 T̄ ψ 1 (τa )ψ 1 (τb )ψ 2 (τc )ψ 2 (τd ) |I

  

with two copies of the twist boundary conditions. On the other hand, due to SO(N ) global
symmetry and large N limit
D E
I|e−βH/2 T̄ [G(τa , τb )G(τc , τd )] |I
D  E
= I|e−βH/2 T̄ ψ 1 (τa )ψ 1 (τb )ψ 2 (τc )ψ 2 (τd ) |I + O(N −1 )

(4.5)

where we suppressed the insertion of e±µV for notational simplicity. In the large N limit,
the leading order contribution to I|e−βH/2 T̄ [G(τa , τb )G(τc , τd )] |I is disconnected, namely



the product of two correlation functions G(τa , τb )G(τc , τd ).

Now we are readily write down the contribution to ρij from this type of OTOC. In
ρ11 , we have
D E
{ψl1 (−t0 ), [ψl2 (−t0 ), U † ψr1 (t)ψr2 (t)U]}
1 XD i E
≈ 2 {ψl (−t0 ), [ψlj (−t0 ), U † ψri (t)ψrj (t)U]}
N
i,j
D ED E
≈ − {ψl1 (−t0 ), U † ψr1 (t)U} {ψl1 (−t0 ), U † ψr1 (t)U}
≈ − K(t, t0 )2 (4.6)

– 22 –
 IRT /(2log2)
1.0

0.8

0.6

0.4

0.2


-1.0 -0.5 0.5 1.0

Figure 5: The plot of IRT /(2 log 2) as a function of K. IRT reaches its maximal value only
when K = ±1.

where in second line we used the identity (4.5) and the fact that ψr1 (t)ψr1 (t) = 1/2, and in
third line we used factorization at large N . Similarly, the correlation function appearing
in ρ14 is given by
D E 1
{ψl2 (−t0 ), ψl1 (−t0 )U † ψr2 (t)Uψl1 (−t0 )} ≈ − K(t, t0 ) (4.7)
2
From these results, we obtain
1
ρ11 = (1 + K(t, t0 )2 ), ρ14 = K(t, t0 ) (4.8)
2
and other ρij via (2.64) to (2.66). From (4.3) we know that K (short for K(t, t0 )) is a real
valued function, which is consistent with the fact that ρ is hermitian. It follows that the
mutual information between R and T is
1
(K − 1)2 log(K − 1)2 + (K + 1)2 log(K + 1)2 + 2(1 − K2 ) log(1 − K2 )

IRT = (4.9)
4
We plot the dependence of K in Fig. 5. We see that IRT is close to maximal when K
approaches to ±1.

4.2 Semiclassical limit

Our analysis relied on the leading 1/N description of the SYK model in terms of classical
G and Σ collective fields. The interaction (2.50) coupled all N fermions ψli with ψri with
a coefficient which we took to scale as µ/(qN ) for fixed µ. We further specialized to the
large q limit, although q  N . In other words, at each order in the 1/q expansion we kept
only the leading effects in 1/N .
Due to the instantaneous nature of the interaction, we treated it exactly in terms of
the fundamental Majoranas, which resulted in particular twist boundary conditions for the
collective fields across the interaction time. In this way, the 1/N effect arising from the
Schwarzian effective action that dominates at large βJ is automatically resummed into
a closed form by the full solution with twist boundary conditions, as in [27]. Moreover,
our analysis includes effects beyond the Schwarzian approximation, which are important

– 23 –
for understanding how the operations involving the fundamental Majoranas glue into the
gravitational dual near the boundary, as discussed below. We will see that those “stringy”
corrections are crucial for regulating the pole in the gravity transmission amplitude in
the probe approximation and are thus very important for finding the effectiveness of the
teleportation protocol. It will not require going to higher orders in the 1/N expansion.
The scrambling time required for traversability is associated to the limit of large N
0
with eω(t+t ) /N fixed. Then K becomes
 
ω 2/q
K = −=  (4.10)
 
h i2/q 
Jµ ω(t+t0 )
J 2/q cosh ω∆t + i 4ωN e e−iωβ/2

In the strong coupling or low temperature limit, βJ  1, the frequency becomes

π 2π
ω = J cos ωβ/2 =⇒ ω = − 2 + O(1/J 2 ) (4.11)
β β J
Plugging this into K, we get
" #2/q 
π/(βJ )
K ≈ −=  J µβ π(t+t0 )/β iπ/(βJ )
 (4.12)
cosh π∆t/β + 4πN e e

which has the maximal Lyapunov exponent of 2π/β [30].

A quick conclusion from (4.12) is that ignoring the numerator, if J , N/J → ∞, a
negative µ will lead to a pole for large enough t + t0 . To compare with gravity result
explicitly, we can define

4∆ βJ
g = −µ, ∆ = 1/q, GN = (4.13)
π∆N
and the first term in (4.12) now becomes

π 2∆
 
1
2∆ (4.14)
βJ

∆g
cosh π∆t/β − G eπ(t+t0 )/β
22∆+2 N

which exactly matches with the semiclassical gravity result of [2] (associated to traversal
of a probe signal, ignoring its backreaction). Such a pole indicates that the bulk geometry
is a traversable wormhole in which there exist timelike geodesics connecting the l system
at −t0 with the r system at t. The negative sign required for µ is also consistent with [16],
in which a large coupling µ that remains on for all time will result in thermofield double
state becoming an eternal traversable wormhole.
An actual pole in the result would violate unitarity, since the fermion number operator
must have an expectation value between 0 and 1. This is consistent with our result, as due
to the existence of numerator, the value of K does not blow up. This comes from the UV
physics of SYK model and cannot be seen in the low energy gravity theory. In the large J
limit, the numerator 1/(βJ ) and the exponential eiπ/(βJ ) perfectly balance to give a finite
maximal value of |K|.

– 24 –
 q 4 8 12 16 20
max(IRT )/(2 log 2) 0.540 0.201 0.105 0.066 0.045

Table 1: Maximal IRT for first a few q’s.

This 1/(βJ ) effect comes from a correction to maximal Lyapunov exponent, thus we
may refer it as a “stringy” effect in the putative dual bulk theory of SYK. For evaluating
the effectiveness of the teleportation protocol, this stringy effect is crucial, as it determines
the actual maximal value of |K|. The nontrivial exponential and the 1/βJ correction to
Lyapunov exponent are also related to the near coherent scrambling discussed in [31].
At weak coupling or high temperature, βJ  1, the frequency is given by ω ≈ J , and
K becomes " #2/q 
1
K ≈ −=  iµ J (t+t0 )
 (4.15)
cosh J ∆t + 4N e

As t + t0 grows, K also has a peak, but it is not as high and sharp as in the low temperature
case. In particular, the denominator then never leads to a pole. This is unsurprising
because at high temperatures, the full stringy dual of the SYK model is unknown, and is
certainly not semi-classical gravity. It is also worth noting that the Lyapunov exponent
now is 2J , as expected [32]. In this paper we are mainly interested in the low temperature
limit and leave discussion of high temperature to Sec. 5.1.
To see the location and value of the peak of K in (4.12), it is useful to define parameters
−µ 2πt/β −µ 2πt0 /β
a ≡ π/(βJ ), x ≡ e , y≡ e (4.16)
4N 4N
in terms of which (4.12) can be written as
1/q
4a4 xy

K =−
4x2 y 2 + a2 (x + y)2 − 4axy(x + y) cos a
" #
2 a(x + y) − 2xy cos a
× sin arccos p (4.17)
q 4x2 y 2 + a2 (x + y)2 − 4axy(x + y) cos a

To find the minimal value of K (note that K < 0), we extremize it with respect to x and
y. Expanding in small a, the solution is
2π
x = y = a + a2 cot + O(a3 ) (4.18)
q+2
which defines a time t = t0 = t0 by (4.16). Using this solution, we find that the minimal
value of K around a ≈ 0 is achieved when J → ∞, a → 0. In this limit, the minimal value
of K is
2π 1+2/q
 
Kmin = − sin (4.19)
q+2
From this expression and (4.9), we see that |Kmin | and max(IRT ) decrease as q becomes
larger. This is reasonable as K is the imaginary part of the first term in (4.12) and q

– 25 –
controls its phase, which decreases as 1/q. For large q the scaling is
max(IRT ) 2π 2
Kmin ∼ −2π/q, ∼ 2 (4.20)
2 log 2 q log 2
We present the maximal values of IRT for a several values of q in Table 1.
With this protocol, the large q limit does not lead to perfect teleportation. However,
the mutual information does grow in time from exponentially small values at early times
to an order 1 number after the scrambling time. By examining the density matrix in more
detail one can determine whether quantum entanglement between R and T is generated,
or if only classical correlation appears. In Sec. 5.4 we show that the resulting density
matrix is inseparable. Thus we could then imagine taking an O(1) number of copies of
SYK model and performing entanglement distillation [33–35], which would increase the
fidelity of teleportation.
It turns out that there is a more elegant way of improving the protocol. In Sec. 4.3 we
shown that applying the initial SWAP operation to a product of order q fermions achieves
almost fidelity of teleportion.
We plot various features of the time dependence of the transmitted signal in Fig. 6.
The final signal, K, is plotted as a function of time, t, for several different initial injection
times −t0 into the left SYK system in Fig. 6. Characterizing the arrival time of the signal
by the location of peak, we see there are two distinct cases. If the injection time is not
sufficiently early (in Fig. 6a), the peaks move to later times when we have earlier injection.
In other words, the time ordering of the signal is inverted. On the other hand, if the
injection time is early enough (in Fig. 6b), the peaks move to earlier times when we have
earlier injection. That means that the time ordering of the signal is preserved.
The latter behavior is what is expected for travel through a traversable wormhole in
semiclassical gravity. Denote the transition time between the two behaviors as t0 = ts .
From Fig. 6a and Fig. 6b, it is clear that the signal significantly strengthens when t0 > ts .
The range t0 > ts should be thought of as the semiclassically traversable regime. Outside of
this semiclassical regime, the time ordering is reversed due to the pattern of entanglement
in the thermofield double state, which is consistent with the analysis of [26].
We can also find the peak arrival time, t for which ∂t K = 0, as a function of the
injection time −t0 . The arrival time t(t0 ) achieves its maximum when t0 = ts . As an
example, for large J we plot this function in terms of exponentiated variables x(y) for
various values of q in Fig. 6c. It is clear that there exists a transition point ts where the
time ordering of the signal flips. For small a (large J ), this transition point in terms of
the variables x and y can be calculated as a series in a:
q + 2 sin2 qθ/2
r
xs = + O(a),
q sin θ
(4.21)
1 a2
ys = a − p + O(a3 )
2 2 q(q + 2) sin θ
where θ is the first positive solution to the transcendental equation
√
qθ q sin θ
tan =√ √ (4.22)
2 q + 2 − q cos θ

– 26 –
  
t t
2 4 6 8 10 12 14 7 8 9 10
-0.0001
-0.2
-0.0002
-0.0003 -0.4
-0.0004
-0.0005 -0.6
-0.0006
-0.8
0
(a) t < ts (b) t0 > ts

2.0

1.5

1.0

0.5

y
0.01 0.02 0.03 0.04 0.05
(c)

Figure 6: The plot of K as a function of arrival time t for different injection time −t0
around transition time −ts . Blue, yellow, green and red correspond to increasing t0 (earlier
injection). (a) shows t0 < ts where peaks move to right and have opposite ordering of
signal, and (b) shows t0 > ts , where peaks move to left and have same ordering of signal.
(c) shows the peak location x(y) for q = 4, 8, 12 (blue, yellow and green). It is clear that
there exists transition point ts at which the time ordering of signal flips. The dashed line
is y = a/2.

2.331
whose solution at large q is θ ≈ q . The transition time ts is defined in terms of ys by

β β 4N
ts ≡ T0 + log ys , T0 = log (4.23)
2π 2π −µ

which is the latest injection time −t0 = −ts for which the signal preserves time ordering.
Similarly, tmax is defined via xs by
β
tmax ≡ T0 + log xs (4.24)
2π
which is the latest arrival time t = tmax after which K has no peak.
Comparing (4.18) with (4.21), we find that the minimal value of K occurs with injection
time −t0 , which is before the transition time. This matches our observation in Fig. 6, and
Kmin in (4.19) is in the semiclassical regime. Moreover, we can plug the value of transition
time (4.21) into K and find K ∼ a1/q in the small a limit. This means that little signal
gets through if injected at or after the transition time, in the low temperature limit. Note
that a ∼ 1/βJ controls “stringy” effects in the bulk. Physically, injections later than

– 27 –
transition time with inverse timing ordering classically do not get through the wormhole at
all and would die in the singularity. However, injections close to the transition time travel
close enough to the (ultimate) horizon that fluctuations in the effective metric determine
whether they get through. Such fluctuations are due to “stringy” effect as indicated by the
|K| peak value being suppressed by (βJ )−1/q .
There is another time scale appears Fig. 6c in the limit y → ∞, so the qubit is injected
very early. It corresponds to the lower bound xmin of x that remains in the semiclassical
regime. we calculate this as a series in a by taking the equation ∂t K = 0 in the t0 → ∞
limit
1 1 2π
xb = a + a2 cot + O(a3 ) (4.25)
2 2 q+2
This defines the earliest semiclassical arrival time
β
tmin ≡ T0 + log xb (4.26)
2π
The physical meaning of tmin is obvious. Any signal arriving semiclassically via the
traversable wormhole cannot arrive earlier than this time. It is reasonable to define the
apparent duration d of the throat of the traversable wormhole, as measured in boundary
time on the right, by
β
d = tmax − tmin ≈ log βJ (4.27)
2π
It is interesting that the duration of the wormhole does not depends on µ and becomes
divergent in the large J limit. These time scales and the two regimes of time orderings of
the signal are shown in Fig. 7.
After the peak, Fig. 6a and Fig. 6b show that K has an exponential decay. This decay
rate is easily seen from (4.12)

K ∼ e−∆(2πt/β) for t large (4.28)

where ∆ = 1/q is the conformal weight of the Majorana fermion ψ. Thus the signal
experiences a thermalization process after transmission to the r system.
We make one final consistency check of our solution in the large N limit. In (3.52), if
we take τa → 3β/4 − it0 and τb → β/4 + it, we get
ω2

σ iµ
e = 2 cosh ω∆t − (cosh ω∆t + i tan ωβ/2 sinh ω∆t)
J N
2
−2
µ tan ωβ/2 0 0
− (i cosh ωt − tan ωβ/2 sinh ωt ) sinh ωt (4.29)
N2
which corresponds to the real correlation function
D E
tfd|U † iψr (t0 )ψl (t)U|tfd (4.30)
0
One can see that the denominator in (4.29) is real when ∆t and eω(t+t /N are order 1,
so that it simply reduces to the first term at leading order in the 1/N expansion. The
imaginary part that appears at higher order cannot be trusted, as then loop corrections
must also be included.

– 28 –
Figure 7: Three time scales of K. We define the arrival time as the peak of K. The signal
emitted from −ts arrives at maximal time tmax and that from −∞ arrives at minimal time
tmin . We define t0 > ts as semiclassical regime because the ordering of signal is preserved
(solid orange arrow). For t0 < ts , the ordering of signal is reversed (dashed orange arrow).

4.3 Improved protocol with multiple fermion SWAP

We found in (4.19) that |Kmin | monotonically decreases with increasing q > 2. It is natural
to expect that by encoding the teleported qubit in multiple fermions, the fidelity of the
protocol can be improved. Noting that the q dependence of the exponent in (4.12) comes
from the 1/q conformal weight of ψ, we will consider making use of heavier operators made
out of a product of several fermions.
To be more precise, we consider the following operators in each SYK system, where p
is an odd number.
2p−1
Ψ1l,r = 2(p−1)/2 ip(p−1)/2 ψl,r
1 3
ψl,r · · · ψl,r , (4.31)
2p
Ψ2l,r =2 (p−1)/2 p(p−1)/2
i 2
ψl,r 4
ψl,r · · · ψl,r (4.32)

which satisfy
{Ψia , Ψjb } = δ ij δab , i, j = 1, 2, a, b = l, r (4.33)
Therefore, we can use Ψi to construct Dirac fermions χ and χ† as in Sec. 2.1
1 1
χl,r = √ (Ψ1l,r + iΨ2l,r ), χ†l,r = √ (Ψ1l,r − iΨ2l,r ) (4.34)
2 2
which implies {χl,r , χ†l,r } = 1. As the Ψi obey exactly the same algebra as a single Majorana
fermion when acting on themselves, the density matrix elements are the same as (2.64)-
(2.66) with the replacement of ψ by Ψ.
We take p to be a fixed number that does not scale with N , so that the correlation
functions involving Ψi will factorize into products of correlations functions of the component
Majonana fermions (with the usual twist boundary conditions for each component). It
follows that the mutual information formula (4.9) for IRT still holds but with K → Kp
given by
D E hD Ep D Ep i
Kp = {Ψl (−t0 ), U † Ψr (t)U} = 2p−1 ψl (−t0 )U † ψr (t)U + U † ψr (t)Uψl (−t0 )
(4.35)

– 29 –
where we omitted the superscript of Ψ. For low temperatures, its value is
" #2p/q 
π/(βJ )
Kp ≈ ip+1 =   (4.36)
cosh π∆t/β + J µβ π(t+t0 )/β iπ/(βJ )
4πN e e

where ip = ±i is pure imaginary as p is an odd number. All of our previous analysis applies
with a rescaled q → q/p. This is reasonable as Ψi consists of p Majorana fermions and
thus has conformal weight p/q.
We are interested in the p value that leads to the highest fidelity teleportation. Odd p
guarantees that (4.35) is pure imaginary, so this means maximizing |Kp |. Using (4.19) we
have  1+2p/q
2π
|Kp |max = sin (4.37)
q/p + 2
where it reaches the maximum value of 1 at p/q = 2. This is perfect fidelity teleportation.
However, as we assumed that q ∈ 4Z and p is odd, we can only choose p = q/2 ± 1. Of
these, p = q/2 + 1 will give the larger |Kp |max . Therefore the improved protocol gives

(q + 2)π 2+2/q
 
|K|max ≡ |Kq/2+1 |max = sin (4.38)
2(q + 1)

which approaches 1 when q → ∞. Expanding at large q, we can see how close the protocol
is to perfect fidelity

π 2 max(IRT ) π2 log q
|K|max ≈ 1 − 2
, ≈ 1 − · 2 (4.39)
4q 2 log 2 4 log 2 q
For high temperatures, the effect of using multiple fermions is similarly a replacement
q → q/p in the previous result (4.15). This leads to a maximal value of |K| given by
 1+2p/q
π
|K|max = sin (4.40)
q/p + 2

which approaches to 1 as p → ∞. We expand it at large p and find

π 2 q max(IRT ) π2 q log p/q

|K|max ≈ 1 − , ≈1− · (4.41)
16p 2 log 2 32 log 2 p
With this improved protocol, we can encode our qubit into the composite fermion field
of q/2 + 1 Majorana fermions in the low temperature regime or into p Majorana fermions
with p  q in the high temperature regime to achieve almost perfect teleportation. The
first case has a dual interpretation of a particle passing through the traversable wormhole.
The high temperature limit of the SYK model must have a very stringy bulk, and it would
be interesting to understand the interpretation of our result in that regime.

– 30 –
5 Discussion

5.1 Interference regime

When an excitation is sent into the wormhole at very early times, its collision with the
negative energy squeezed state is highly trans-Planckian and the backreaction is large. It
was shown in [2, 26] that nevertheless the left-right correlation function K remains nonzero
due to an interference effect. We now examine whether this interference effect can lead to
teleportation with our protocol. Note that the above large q solution is related to semi-
classical gravity plus “stringy” effects and does not include the interference regime, which
goes beyond the leading 1/N contributions. Therefore, we have to discuss it separately.
The interference effect relies on the facts that time-ordered correlation functions fac-
torize and out-of-time-ordered correlation functions vanish due to the nature of quantum
chaos. Basically, we have
eiµV |tfdi ≈ eiµhV i |tfdi ,
j1 js j1 js
(5.1)
eiµV ψl,r · · · ψl,r |tfdi ≈ ψl,r · · · ψl,r |tfdi
where hV i = i hψl (0)ψr (0)i /q is an order 1 number. Note that the large q solution above
does not capture the interference regime because K vanishes for large t and t0 in (4.3)
whereas the interference effect implies that it should be 2<(eiµhV i hψl (−t0 )ψr (t)i). Using
(5.1) we find that
D E µ hV i
1 1 2
{ψl1 , [ψl2 , U † ψr1 ψr2 U]} → 4 sin2 ψl ψr (5.2)
D E 2
{ψl1 , U † ψr1 U} → 2i sin µ hV i ψl1 ψr1


(5.3)
D E
{ψl2 , ψl1 U † ψr2 Uψl1 } → 0 (5.4)

which leads to
1 µ hV i
1 1 2
ρ11 = − 2 sin2 ψl ψr (5.5)
2 
12 1 
|ρ14 | = sin µ hV i  ψl ψr  (5.6)

1 1
where h·i is short for htfd| · |tfdi. Note that ψl ψr is pure imaginary because the cor-
relation function between ψl and iψr is real. Moreover, | ψl1 ψr1 | is bounded by 1/2 via



1 1
Schwarzian inequality. For different values of | ψl ψr |, we plot IRT as a function of µ hV i
in Fig. 8.
In the low temperature limit, | ψl1 ψr1 | ∼ hV i scales as 21 (π/βJ )2/q which is a small



number (much smaller than order 1). From Fig. 8 we see that IRT is quite small and not
sufficient for successful teleportation. More precisely, one can show that

IRT ∼ (βJ )−8/q (5.7)

1 1
On the other hand, in the high temperature limit, ψl ψr → i/2, which leads to an order 1
magnitude for IRT . In this case, IRT is bounded by log 2. Although the mutual information
does not reach maximality, for appropriate choices of µ some quantum information is
teleported via the interference effect, as we will discuss in Sec. 5.4.

– 31 –
 IRT /(2log2)
0.5

0.4

0.3

0.2

0.1

μ〈V〉
1 2 3 4 5 6

Figure 8: Mutual information IRT as a function of µ hV i for different | ψl1 ψr1 |. Blue,



yellow, green and red represent for | ψl1 ψr1 | = 0.2, 0.3, 0.4, 0.5 respectively.



5.2 Partial coupling

In this section we consider a simple generalization of the protocol in which only a fraction
of the Majorana fermions are coupled in the interaction Hamiltonian. This is also required
for the coupling being interpreted as a classical channel as discussed in Sec. 2.1. It is
particularly interesting to consider measuring the correlation function of fermions that do
not appear in the interaction. The quantum chaotic nature of the SYK Hamiltonian implies
that there need be no relation between the fermions that are coupled with the fermions
that are transmitted, as expected in the gravity description.
We consider coupling only the first αN Majorana fermions7
αN
iµ X j j
Hint = −µṼ = − ψl ψr (5.8)
qαN
j=1

and we calculate the correlation function for the last (1 − α)N Majorana fermions
X D E
0
I|e−βH/2 T̄ [e−µṼ (β/4+) eµṼ (β/4−) ψ j (τa )ψ j (τb )]|I (5.9)
j

where 0 means summing j from αN + 1 to N . It is clear that ψ j can freely move across
P

Ṽ for j = αN + 1, · · · , N , so they do not obey a twisted boundary condition. One might

naively conclude that such correlation functions do not have any of the nontrivial features
of Gµ in the various subregions in Fig. 4 that we found before, in particular that there
would be no exponential growth in subregion c after continuation to Lorentz signature.
However, this expectation is incorrect. Indeed, from the general analysis of the SYK
model, the low energy effective action is the Schwarzian that universally couples to all
fermions [15]. Thus the OTOC between different species of fermions should have exponen-
tial growth at order 1/N . Moreover, traversable wormholes should exist for generic coupled
operators based on analysis of both the gravity [1] and CFT [26] sides.
7
Here we assume αN is an integer of order N .

– 32 –
 The resolution is that we need to treat the effective action (2.47) more carefully. As
we separated the fermions into two groups, we need to define two different Σ and G fields
αN
0 1 X i
G1 (τ, τ ) = ψ (τ )ψ i (τ 0 ) (5.10)
αN
i=1
N
1 X
G2 (τ, τ 0 ) = ψ i (τ )ψ i (τ 0 ) (5.11)
(1 − α)N
i=αN +1

The instantaneous coupling adds new terms to the action

µ
δS = [G1 (3β/4 + , β/4 − ) − G1 (β/4 − , 3β/4 + )
2q
+G1 (β/4 + , 3β/4 − ) − G1 (3β/4 − , β/4 + )] (5.12)

Therefore the effective action becomes

det(∂τ − Σ2 )]α−1 J2
 ZZ  
N q
S= log + αΣ1 G1 + (1 − α)Σ2 G2 − (αG1 + (1 − α)G2 ) + δS
2 det(∂τ − Σ1 )α q
(5.13)

which leads to the following equation of motion

Gi =[∂τ − Σi ]−1 , (5.14)

µ
Σ1 =J 2 (αG1 + (1 − α)G2 )q−1 + δ̃, (5.15)
qαN
Σ2 =J 2 (αG1 + (1 − α)G2 )q−1 , (5.16)
δ̃ =δτ,β/4− δτ 0 ,3β/4+ − δτ,3β/4+ δτ 0 ,β/4− − δτ,β/4+ δτ 0 ,3β/4− + δτ,3β/4− δτ 0 ,β/4+ (5.17)

where i = 1, 2 and δa,b is short for δ(a − b). As expected Σ2 and G2 does not directly have
a delta function in their equations of motion. However, the two parts interact with each
other and the twist boundary condition for G1 affect G2 accordingly. As discussed in Sec.
2.2, the equations of motion only captures the linearized version of the twist boundary
condition. Luckily, in the large q solution, this is enough. To be precise, let us expand

[Gi ] ≈ [G0 ] + [G0 ][Σi ][G0 ] (5.18)

Defining G+ = αG1 + (1 − α)G2 and G− = G1 − G2 , the equations of motion become

J2 µ
[G+ ] =[G0 ] + [G0 ][(2G+ )q−1 ][G0 ] + [G0 ][δ̃][G0 ] (5.19)
q qN
µ
[G− ] = [G0 ][δ̃][G0 ] (5.20)
qαN
The first equation is exactly the same as the one for Gµ in our earlier analysis, and we can
solve it with the same ansatz
G+ = G0 eσ+ /(q−1) (5.21)

– 33 –
where σ+ obeys the boundary conditions (3.38)-(3.40). The second equation can be eval-
uated explicitly
µ X
G− = [±G0 (τ, β/4 ∓ )G0 (3β/4 ± , τ 0 ) ∓ G0 (τ, 3β/4 ± )G0 (β/4 ∓ , τ 0 )] (5.22)
qαN ±

In the fundamental region, the configuration is very simple

µ
G− = − in c, G− = 0 in others (5.23)
2qαN
Since G1 = G+ + (1 − α)G− and G2 = G+ − αG− , they are identical in all subregions
of fundamental region except c. In subregion c, both G1 and G2 have exponential growth
after continuation to Lorentz signature as G+ does. However, on the boundary of c, the
boundary condition for σ+ is
µ
σ+ (inside c) = σ+ (outside c) − (5.24)
N
To leading order in 1/q, this shift perfectly cancels the constant shift in G− ,
1 1 µ
G2 (inside c) ≈ (1 + σ+ (inside c)) +
2 q 2qN
1 1
= (1 + σ+ (outside c)) ≈ G2 (outside c) (5.25)
2 q
which shows that G2 is continuous. This analysis clearly shows the crucial feature of
traversable wormhole teleportation that the coupled qubits between two systems can be
chosen independently from the teleported qubits.
Note that the number of coupled qubits must be of order N to guarantee that the path
integral over the effective action (5.13) is dominated by the classical solution. Otherwise,
one must treat all αN ∼ O(1) Majorana Fermions with full quantum corrections, which is
beyond the regime of validity of the calculations in this paper.

5.3 Teleportation of multiple qubits

If we want to teleport n qubits together, we can simply prepare R = ∪i Ri , T = ∪i Ti and
Q = ∪i Qi for i = 1, · · · , n and apply n SWAP operators associated to n different (Dirac)
fermions in the SYK model. As these SWAP operators have different fermion indices, they
do not talk to each other, due to the SO(N ) approximate symmetry. Therefore, correlation
functions with twist boundary conditions factorize into independent channels in the large
N limit. It follows that the reduced density matrix between R and T is tensor product of
n identical matrices, each of which is given as before (2.13). The caveat in this argument
is that the number of teleported qubit n must be o(N ), otherwise the factorization of the
correlation functions at large N breaks down. Since o(N ) < O(N ), this is consistent with
the teleportation information bound: 2 classical bits can only teleport at most one qubit.
This is also consistent with the upper bound of total number of teleported qubit from
gravitational computation [36].
Although the traversable wormhole teleportation does not teleport the maximal num-
ber of qubits, it does have an advantage that the qubits are strongly error protected [37].

– 34 –
Indeed, as we can choose arbitrary O(N ) Majorana fermions in the interaction V , it does
not matter what happens to other uncoupled ones. In the quantum information language,
when some qubits are corrupted or erased, we can still measure other intact qubits in the l
system to complete the teleportation protocol. Holographically, the dual picture is a qubit
traveling through a traversable wormhole. As the location of qubit is in the deep interior, it
is not sensitive to local boundary changes, which correspond to simple (O(1) qubit) errors.

5.4 Classical simulation and separability

As we have seen in Sec. 4.3, an appropriate choice of SWAP operator leads to almost
perfect fidelity in the large q limit. However, in the interference regime and for other
choices of the SWAP in the large q solution, the mutual information can still be O(1)
although submaximal, so there is not perfect teleportation. Here we examine in more
detail whether teleportation in those cases really communicates quantum information that
cannot be simulated by a purely classical communication protocol. In other words, we need
to investigate if ρRT is separable.
Generally, a density matrix on two systems ρRT is separable if it can be written in the
form of sum of tensor products of density matrices in each subsystem,
X X
ρRT = pi ρiR ⊗ ρiT , pi = 1, pi > 0, ∀i (5.26)
i i
Here pi can be understood as the classical probability for each tensor product state. By
definition, it is clear that the mutual information is purely classical if the density matrix
ρRT is separable. In such a case, we could achieve the same result with classical simulation,
and our quantum teleportation protocol would have failed.
For a general bipartite density matrix, determining if it is separable is an NP-hard
problem [38, 39]. However, for our pair of single qubit systems, the sufficient and nec-
essary condition is the Peres-Horodecki criterion [40, 41], which states that if the partial
transposed density matrix has a negative eigenvalue, then the density matrix is quantum
entangled. By (2.64)-(2.66), the partial transpose of the density matrix ρTR is simply
obtained by exchanging ρ14 with ρ23 .
In our large q solution, the eigenvalues of ρTR are
1 1
(1 + K2 ) (multiplicity 2), (1 ± 2K − K2 ) (5.27)
4 4
The first expression is always positive and the second one could be negative when |K| >
√
2 − 1 ≈ 0.414. For low temperatures with the improved protocol, the maximum value
of |K| is given by (4.37), which satisfies this bound when 0.094q < p < 1.260p. For high
temperatures with the improved protocol, the maximum value of |K| is given by (4.40),
which satisfies the bound when p > 0.336q. In those cases, R and T have some quantum
entanglement.
In the interference regime, the eigenvalues of ρTR are
1
(1 + 2U 2 − 2U 2 cos µ hV i) (multiplicity 2), (5.28)
4
1
(1 − 2U 2 + 2U 2 cos µ hV i ± 2U sin µ hV i), (5.29)
4

– 35 –
where U ≡ −i hψl (−t0 )ψr (t)i ∈ (0, 1/2). The first expression is always positive and the
second one could be negative when
p
1 − sin µ hV i + (1 − cos µ hV i)(3 + 2 cos µ hV i)
≥U > (5.30)
2 2(1 − cos µ hV i)

It is clear that in the high temperature limit, U = |V | = 1/2 and above inequality is
satisfied when 2.294 + 4kπ < µ < 7.131 + 4kπ for k ∈ Z. In this case, there is teleportation
of some quantum information in the interference regime.

6 Conclusion

In this paper, we proposed a simple teleportation protocol using a pair of SYK systems.
The protocol requires the two systems prepared in the highly entangled thermofield double
state. There are three steps:

1. Apply a SWAP operation in one system to insert a qubit at early (negative) time
t = −tin ;

2. Evolve the state to t = 0 and apply a weak coupling between two SYKs;

3. Evolve the state to t ≈ tin and apply a SWAP to extract the qubit.

We calculated the mutual information IRT between a reference system and the final
qubit to confirm the effectiveness of this protocol. The protocol has a clear holographic
interpretation in which the teleported qubit goes through a traversable wormhole in a
nearly AdS2 spacetime. At low temperature, the result matches the semiclassical gravity
analysis in two aspects: first, we see a high peak in IRT at a specific time that is related to
the geodesics through the traversable wormhole; second, there is a range of time in which
time ordering of the signal is preserved. We take these as signs that the information travels
smoothly through the throat of a traversable wormhole. As the temperature increases, we
identify some “stringy” corrections to correlation functions that will soften the peak in
IRT .
Our protocol is quite similar to standard teleportation in the sense that the weak
coupling between two SYKs can be implemented as a measurement in one system and
a corresponding unitary in the other. However, the measurement in our protocol is in
the computational basis thanks to scrambling, and the teleported qubit is highly error
protected. This protocol allows teleportation of o(N ) qubits simultaneously, at the cost of
communication of O(N ) bits of classical information, which obeys the information bound
on quantum teleportation.
The protocol can be improved to approach perfect fidelity by an appropriate choice
of SWAP operator. For large q SYK at low temperatures, using a SWAP of the qubit
into q/2 + 1 Majorana fermions, we find that the mutual information IRT tends to 1 when
q → ∞. Similarly, at high temperature, SWAP of the qubit into p Majorana fermions with
p  q also results in mutual information close to 1.

– 36 –
 At very late times, we also analyzed the teleportation effect in the fully scrambled
interference regime. It turns out that the interference effect makes teleportation possible
only at high temperature rather than low temperature with our simple protocol. This cor-
responds to very late time black holes, for which Kitaev and Yoshida proposed an algorithm
[11] involving complicated operators to decode the scrambled qubit. Our protocol has a
much simpler structure and works for both around the scrambling time at all temperatures,
and for very late times at high temperature.

Acknowledgments

We thank Xun Gao, Yingfei Gu, Emil Khabiboulline, Thomas Schuster, Maria Spiropulu
and Norman Yao for helpful and stimulating discussions. PG would like to thank Tsung-
Dao Lee Institute of Shanghai Jiaotong University for hospitality during the period of this
project. PG and DLJ are both supported by DOE award DE-SC0019219. Besides, PG
is supported by the US Department of Energy grants DE-SC0018944 and DE-SC0019127,
and also the Simons foundation as a member of the It from Qubit collaboration.

A Solution of σ

In this appendix, we provide the details of how to determine σ. Let us assume that

−µ ãµ + b̃µ f0−µ

g0µ = SLµ (f0 ) ≡
f (A.1)
c̃µ + d˜µ f0−µ
which leads to
f µ (f −µ )) ' (SL
(f0µ , SL f −1 (f −µ )),
f −µ (f µ ), f −µ ) ' (f µ , SL
0 0 0 0 −µ 0
˜ 2
f −µ (f ) ≡ d−µ + b̃−µ J f
SL
−1
(A.2)
J 2 (c̃−µ + ã−µ J 2 f )
where the first ' is due to the symmetry (3.43) and the second ' is due to the global
SL(2) symmetry (3.29) of the large q solution. When µ = 0, we see from (3.36) that SL f
˜
is the identity, from which we can fix b̃µ c̃µ − ãµ dµ = 1 because we want the solution to
smoothly reduce to the µ = 0 case. It follows that

ãµ + b̃µ f0−µ d˜−µ + b̃−µ J 2 f0−µ

=
c̃µ + d˜µ f0−µ J 2 (c̃−µ + ã−µ J 2 f0−µ )
=⇒ ãµ J 2 − d˜−µ = b̃µ − b̃−µ = c̃µ − c̃−µ = 0 (A.3)

Similarly, f0µ must be related to f0∗

µ
by an SL(2) transformation, explicitly
µ
αµ + βµ f0∗
f0µ = slµ (f0∗
µ
)≡ µ (A.4)
γµ + δµ f0∗
This SL(2) squares to the identity because f0µ = slµ (f0∗
µ
) = slµ (slµ (f0µ )), which implies
that
βµ + γµ = 0, or αµ = 0, δµ = 0 (A.5)

– 37 –
Comparing with (3.36), we see that

− J1 sin 2v + cos 2v · f 1
f∗ = , f= tan(ωτ + v) (A.6)
− cos 2v − J sin 2v · f J
which has nonzero αµ and δµ under τ reflection. Therefore we must take the former choice
in (A.5) and fix βµ2 + αµ δµ = 1.
Since we have two SL(2) symmetries in region b, they need to be consistent with each
other
f µ (f −µ )) ' (f µ , SL
(f0µ , SL f µ (f −µ )) (A.7)
0 0∗ 0∗

In other words, as slµ transforms f0∗µ f µ (f −µ ) to SL

to f0µ , it must transform SL f µ (f −µ )
0∗ 0
accordingly. We can follow similarly to the derivation of (A.2) to get

slµ−1 (SL
f µ (f )) = SL
f µ (sl−µ (f )) (A.8)

which leads to

ãµ βµ + c̃µ δµ /J 2 = b̃µ α−µ − ãµ β−µ , (A.9)

b̃µ βµ + d˜µ δµ /J 2 = ãµ δ−µ + b̃µ β−µ (A.10)
J 2 ãµ αµ − c̃µ βµ = d˜µ α−µ − c̃µ β−µ , (A.11)
J b̃µ αµ − d˜µ βµ = c̃µ δ−µ + d˜µ β−µ
2
(A.12)

which can be solved as

q
−b̃µ β−µ ± b̃2µ − δµ δ−µ /J 2
ãµ = ,
δ−µ
q (A.13)
b̃µ (1 + βµ β−µ ) ∓ (βµ + β−µ ) b̃2µ − δµ δ−µ /J 2
c̃µ = J 2
δµ δ−µ

with b̃µ = b̃−µ , d˜µ = J 2 ã−µ and αµ = (1 − βµ2 )/δµ by definition. These are all constraints
resulting from the symmetries and smoothness requirements.
We can apply a similar analysis to regions a and e. There is only one symmetry in
each of these two regions, so the constraint is much weaker. Taking all these into account,
we assign functions to each subregion in Fig. 4 with only three functions f µ , g1µ and g µ
(here we simplify notation f0µ → f µ ), and four SL(2) transformations SL f µ , SL
c µ , SLµ and
slµ to be determined, where SL c µ and SLµ obey

âµ J 2 − dˆ−µ = b̂µ − b̂−µ = ĉµ − ĉ−µ = āµ J 2 − d¯−µ = b̄µ − b̄−µ = c̄µ − c̄−µ = 0 (A.14)

with b̂µ ĉµ − âµ dˆµ = 1 and b̄µ c̄µ − āµ d¯µ = −1. Here the normalization for SLµ is −1 to
−µ
agree in the µ = 0 case with g1µ = SLµ (g1∗ ) = f in (3.36).
The twist boundary conditions on the four blue segments in Fig. 4 are given by
(3.45)-(3.48). We can use these equations to determine both SLµ and SL c µ based on the
value of f µ and g1 on the boundaries. For notational simplicity, we define f±µ ≡ f µ (τ± ),
µ

– 38 –
 µ
f±∗ ≡ f µ (−τ± ) and similarly for g1µ and their derivatives. Using (A.13) and f−µ we can
µ µ 0
determine g1− and g1− via (3.46). Based on these values, we can determine SL
c µ completely
via (3.47), the constraints (A.14) and the value (A.13). It turns out that

b̂µ = ∓c̃µ , ĉµ = ∓b̃µ , âµ = ±ã−µ , dˆµ = J 2 â−µ (A.15)

As both signs give the same SL c µ , we will choose the lower sign without loss of generality.
µ
To solve for SLµ is similar. The only difference is that it involves f−∗ that is related to f−µ
via slµ . It turns out that one can determine āµ , b̄µ , c̄µ and d¯µ explicitly in terms of b̃µ , β±µ ,
δ±µ and f±µ only. However, the expression is very complicated and we do not reproduce it
here. At this point, the free parameters have been reduced to b̃µ , β±µ , δ±µ .
As argued in Sec. 3.2.3, there are still too many degree of freedom to uniquely fix the
solution. Taking the ansatz (3.51), we immediately get

βµ = β = cos 2v, δµ = δ = −J sin 2v =⇒ ãµ = d˜µ = âµ = dˆµ = 0, c̃µ = b̂µ = 1 (A.16)

On the other hand, we know

 
−µ a + bf
(SLµ (g1∗ ), g1µ ) ' ,f , bc − ad = 1 (A.17)
c + df

for some a, b, c, d. Applying this to the formula for eσ and checking the UV condition, we
find
b = c = cos ϕ, d = −aJ 2 = −J sin ϕ,
ω2 (A.18)
eσ = 2 2
, ω = J cos(ϕ − ωβ/2)
J cos (ωτab + ϕ)
Now we can forget about SLµ and directly impose the symmetry conditions: µ → −µ,
τa,b → −τb,a . This requires that ϕ is invariant under µ → −µ. In order to more simply
satisfy the twist boundary condition, we use the global SL(2) to write
     
a + bf a + bf
,f ' F , G(f ) (A.19)
c + df c + df

where F is given by (3.29)

D µ − Cµ J 2 f
F (f ) = (A.20)
J 2 (−Bµ + Aµ J 2 f )

Let us check twist boundary conditions using (3.51), (A.20) and (A.18):

G(f− ) = f− , G0 (f− ) = e−µ/N , (A.21)

 
  F 0 a+bf−∗
a + bf−∗ c+df−∗
F = f−∗ , = eµ/N (A.22)
c + df−∗ (c + df−∗ )2

These four equations determine the four variables: Aµ , Bµ , Cµ , ϕ, in which

ϕ
cot = cot 2ωτ− or ϕ = 0 (A.23)
2

– 39 –
As we would like to match on to the µ = 0 solution smoothly, we must choose the second
case. It follows that
µ
2f− sinh 2N J 2 f− f−∗ eµ/(2N ) + e−µ/(2N )
Aµ = , B µ = , (A.24)
1 + J 2 f− f−∗ 1 + J 2 f− f−∗
µ
J 2 f− f−∗ e−µ/(2N ) + eµ/(2N ) 2J 2 f−∗ sinh 2N
Cµ = , D µ = (A.25)
1 + J 2 f− f−∗ 1 + J 2 f− f−∗
and UV condition becomes
ω = J cos(ωβ/2) (A.26)
Lastly, we will solve for g µ . This is easier to do in terms of the unshifted τa coordinate.
In the unshifted coordinate,
tan(ω(τ − β/2) + v) − 1 tan ωβ/2 + f
fµ = = J (A.27)
J 1 + J tan ωβ/2 · f
This is an SL(2) transformation. We can simply assume that in the unshifted coordinate
f µ = f , and find the most general g µ that satisfies the UV condition. Then doing a corre-
sponding SL(2) transformation (3.29) to g µ will give the solution in the shifted coordinate.
Let us assume g µ = H(f ) for some functional of f . The UV condition becomes
H 0 (f )f 02
=1 (A.28)
(1 + J 2 H(f )f )2
Note that for our special choice of f , f 0 is related to f by
ω2 ω2
f 02 = sec 4
(ωτ + v) = (1 + J 2 f 2 )2 (A.29)
J2 J2
Plug this back to (A.28) and solve for H as
1 hµ sin(ωτ + v − φ) − e2(ωτ +v) tan φ sin(ωτ + v + φ)
H(f ) = (A.30)
J hµ cos(ωτ + v − φ) − e2(ωτ +v) tan φ cos(ωτ + v + φ)
where ω = J cos φ and hµ is an integration constant. Comparing with (A.26), we have
φ = ωβ/2. Before shifting it back to the new coordinate τa → τa + β/2, we apply (A.27)
to f and an associated SL(2) transformation to g µ :
1
Jtan ωβ/2 − g µ
gµ → (A.31)
−1 − J tan ωβ/2 · g µ
which leads to
1 e2(ωτ +v) tan ωβ/2 sin(ωτ + v) − hµ sin(ωτ + v − ωβ)
gµ = (A.32)
J e2(ωτ +v) tan ωβ/2 cos(ωτ + v) − hµ cos(ωτ + v − ωβ)
Solving twist boundary condition, we find
e2(ωτ+ +v) tan ωβ/2 (e−µ/N − 1) sin ωτ+− sin(ωτ− − v)
hµ =
cos ωβ/2 cos(ωτ− + v) − cos 2ωτ− cos(ωτ+ − ωβ + v) + e−µ/N sin ωτ+− sin(ωτ− − v)
(A.33)
which vanishes when τ+ → τ− . This is a consistency requirement. It is important that we
need a nontrivial hµ for a consistent solution to our problem, even though in the end we
do not need it as we will take the τ+ → τ− limit.

– 40 –
References

[1] P. Gao, D.L. Jafferis and A.C. Wall, Traversable Wormholes via a Double Trace
Deformation, JHEP 12 (2017) 151 [1608.05687].
[2] J. Maldacena, D. Stanford and Z. Yang, Diving into traversable wormholes, Fortsch. Phys.
65 (2017) 1700034 [1704.05333].
[3] J. Maldacena and L. Susskind, Cool horizons for entangled black holes, Fortsch. Phys. 61
(2013) 781 [1306.0533].
[4] J. Maldacena and A. Milekhin, Humanly traversable wormholes, 2008.06618.
[5] J. Maldacena, A. Milekhin and F. Popov, Traversable wormholes in four dimensions,
1807.04726.
[6] S. Fallows and S.F. Ross, Making near-extremal wormholes traversable, JHEP 12 (2020) 044
[2008.07946].
[7] R. Emparan, B. Grado-White, D. Marolf and M. Tomasevic, Multi-mouth Traversable
Wormholes, 2012.07821.
[8] A.A. Balushi, Z. Wang and D. Marolf, Traversability of Multi-Boundary Wormholes,
2012.04635.
[9] L. Susskind and Y. Zhao, Teleportation through the wormhole, Phys. Rev. D 98 (2018)
046016 [1707.04354].
[10] N. Bao, A. Chatwin-Davies, J. Pollack and G.N. Remmen, Traversable Wormholes as
Quantum Channels: Exploring CFT Entanglement Structure and Channel Capacity in
Holography, JHEP 11 (2018) 071 [1808.05963].
[11] B. Yoshida and A. Kitaev, Efficient decoding for the Hayden-Preskill protocol, 1710.03363.
[12] A.R. Brown, H. Gharibyan, S. Leichenauer, H.W. Lin, S. Nezami, G. Salton et al., Quantum
Gravity in the Lab: Teleportation by Size and Traversable Wormholes, 1911.06314.
[13] S. Sachdev and J. Ye, Gapless spin-fluid ground state in a random quantum heisenberg
magnet, Physical Review Letters 70 (1993) 3339.
[14] A. Kitaev, A simple model of quantum holography, in KITP strings seminar and
Entanglement, vol. 12, 2015.
[15] J. Maldacena and D. Stanford, Remarks on the Sachdev-Ye-Kitaev model, Phys. Rev. D 94
(2016) 106002 [1604.07818].
[16] J. Maldacena and X.-L. Qi, Eternal traversable wormhole, 1804.00491.
[17] T. Numasawa, Four coupled SYK models and Nearly AdS2 gravities: Phase Transitions in
Traversable wormholes and in Bra-ket wormholes, 2011.12962.
[18] N. Sorokhaibam, Traversable wormhole without interaction, 2007.07169.
[19] Y. Chen and P. Zhang, Entanglement Entropy of Two Coupled SYK Models and Eternal
Traversable Wormhole, JHEP 07 (2019) 033 [1903.10532].
[20] Y.D. Lensky and X.-L. Qi, Rescuing a black hole in the large-q coupled SYK model,
2012.15798.
[21] J. Maldacena and A. Milekhin, SYK wormhole formation in real time, 1912.03276.

– 41 –
[22] S. Plugge, E. Lantagne-Hurtubise and M. Franz, Revival Dynamics in a Traversable
Wormhole, Phys. Rev. Lett. 124 (2020) 221601 [2003.03914].
[23] X.-L. Qi and P. Zhang, The Coupled SYK model at Finite Temperature, JHEP 05 (2020) 129
[2003.03916].
[24] S.H. Shenker and D. Stanford, Stringy effects in scrambling, JHEP 05 (2015) 132
[1412.6087].
[25] Y. Gu and A. Kitaev, On the relation between the magnitude and exponent of OTOCs, JHEP
02 (2019) 075 [1812.00120].
[26] P. Gao and H. Liu, Regenesis and quantum traversable wormholes, JHEP 10 (2019) 048
[1810.01444].
[27] X.-L. Qi and A. Streicher, Quantum Epidemiology: Operator Growth, Thermal Effects, and
SYK, JHEP 08 (2019) 012 [1810.11958].
[28] A. Eberlein, V. Kasper, S. Sachdev and J. Steinberg, Quantum quench of the
Sachdev-Ye-Kitaev Model, Phys. Rev. B 96 (2017) 205123 [1706.07803].
[29] D.J. Gross and V. Rosenhaus, A Generalization of Sachdev-Ye-Kitaev, JHEP 02 (2017) 093
[1610.01569].
[30] J. Maldacena, S.H. Shenker and D. Stanford, A bound on chaos, JHEP 08 (2016) 106
[1503.01409].
[31] A. Kitaev, Near coherent scrambling, IAS program Quantum Information and Black Holes
57 (2017) .
[32] D.A. Roberts, D. Stanford and A. Streicher, Operator growth in the SYK model, JHEP 06
(2018) 122 [1802.02633].
[33] C.H. Bennett, H.J. Bernstein, S. Popescu and B. Schumacher, Concentrating partial
entanglement by local operations, Phys. Rev. A 53 (1996) 2046 [quant-ph/9511030].
[34] C.H. Bennett, D.P. DiVincenzo, J.A. Smolin and W.K. Wootters, Mixed-state entanglement
and quantum error correction, Physical Review A 54 (1996) 3824.
[35] C.H. Bennett, G. Brassard, S. Popescu, B. Schumacher, J.A. Smolin and W.K. Wootters,
Purification of noisy entanglement and faithful teleportation via noisy channels, Phys. Rev.
Lett. 76 (1996) 722 [quant-ph/9511027].
[36] B. Freivogel, D.A. Galante, D. Nikolakopoulou and A. Rotundo, Traversable wormholes in
AdS and bounds on information transfer, JHEP 01 (2020) 050 [1907.13140].
[37] A. Almheiri, X. Dong and D. Harlow, Bulk Locality and Quantum Error Correction in
AdS/CFT, JHEP 04 (2015) 163 [1411.7041].
[38] L. Gurvits, Classical deterministic complexity of edmonds’ problem and quantum
entanglement, in Proceedings of the Thirty-Fifth Annual ACM Symposium on Theory of
Computing, STOC ’03, (New York, NY, USA), p. 10–19, Association for Computing
Machinery, 2003, DOI.
[39] S. Gharibian, Strong np-hardness of the quantum separability problem, 0810.4507.
[40] A. Peres, Separability criterion for density matrices, Physical Review Letters 77 (1996)
1413–1415.

– 42 –
[41] P. Horodecki, Separability criterion and inseparable mixed states with positive partial
transposition, Physics Letters A 232 (1997) 333–339.

– 43 –